# Source paper archive

- Source title: Automated detection and classification of morphologically similar weevil species using AI for precision pest monitoring
- Author(s): Tukuli, A. R., Perkin, L. C., Raszick, T. J., Shirley, X. A., Cohen, Z. P., Suh, C. P.- C.
- Source landing URL: https://www.biorxiv.org/content/10.64898/2026.11.17.767224v1
- PDF URL: https://www.biorxiv.org/content/10.64898/2026.11.17.767224v1.full.pdf
- PDF SHA-256: 5abf4773593ccd99896fd0f784339f40131f3fa1c096efeb655dc5b9d6b9174c
- Converter: pymupdf4llm 0.0.27
- Limits: Automated PDF text extraction; reading order, equations, tables and formatting may be incomplete or incorrect. Images are omitted, and no OCR was performed. Markup and links are neutralized for safe downloading. Consult the original PDF for the authoritative version.

---

&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.17.767224; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.17.767224)

(which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC
105 and is also made available for use under a CC0 license.

### 1 Automated detection and classification of morphologically similar 2 weevil species using AI for precision pest monitoring


3 _Short title: Automated weevil identification with RRWIT_


4 Adama R. Tukuli &#91;1*&#93;, Lindsey C. Perkin &#91;1*&#93;, Tyler J. Raszick &#91;2&#93;, Xanthe A. Shirley &#91;3&#93;, Zachary P.


5 Cohen &#91;1&#93;, and Charles P.-C. Suh &#91;1&#93;


6 1 USDA-ARS Southern Plains Agricultural Research Center, College Station, Texas, United


7 States of America


8 2 USDA-APHIS-PPQ Science &amp; Technology, Insect Management and Molecular Diagnostics


9 Laboratory, Edinburg, Texas, United States of America


10 3 USDA-APHIS-PPQ College Station, Texas, United States of America


11


12 - Corresponding authors: &#91;adama.tukuli@usda.gov&#93;(mailto:adama.tukuli@usda.gov) (AT), &#91;lindsey.perkin@usda.gov&#93;(mailto:lindsey.perkin@usda.gov) (LP)


13 **Author contributions**


14 **Conceptualization:** AT, LP, ZC, CS, TR, XS; **Designed the study, specimen acquisition and**


15 **verification:** AT, LP, ZC, CS, TR, XS; **Methodology, formal analysis and initial draft:** AT. All


16 authors contributed to writing and reviewing the manuscript.


17


18


19


20


21


22


23


24


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.17.767224; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.17.767224)

(which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC
105 and is also made available for use under a CC0 license.

# 25   Abstract


26 Accurate identification of morphologically similar insects is important for agricultural


27 surveillance but remains challenging when large and balanced image datasets are


28 limited. We developed Rostrum Recognition: A Weevil Identification Tool (RRWIT), a


29 lightweight object-detection framework for distinguishing 14 weevil taxa, including two


30 economically important weevil species, the boll weevil and pecan weevil, using more


31 than 16,000 expert-verified annotated images. RRWIT was developed by transfer


32 learning from a pretrained YOLO11n detector using a stratified data partition with the


33 test set withheld from model and threshold selection. We evaluated this design with


34 matched random-initialization and transfer-learning experiments, quantitative


35 EigenGradCAM attribution analysis, targeted architectural modifications, and a frozen


36 eight-framework benchmark across three independent training seeds. In the cross

37 framework benchmark, RRWIT achieved a mean average precision at an intersection

38 over-union threshold of 0.50 (mAP50) of 0.938 ± 0.005 on held-out test partitions subset,


39 whereas the newer lightweight YOLO26n achieved the highest value, 0.976 ± 0.010,


40 with a similarly compact computational footprint. Post-training 8-bit integer (INT8)


41 quantization reduced the serialized RRWIT model from 10.09 to 2.91 MiB (71.2%) with a


42 small validation mAP50 change from 0.978 to 0.971 but did not accelerate end-to-end


43 inference under the tested central processing unit configuration. RRWIT therefore


44 provides an extensible foundation for automated weevil identification while highlighting


45 the importance of representative biological data, independent external validation,


46 evolving lightweight architectures, and hardware-specific deployment optimization in


47 future development. A publicly accessible RRWIT web application supporting image


48 upload and live image capture for automated weevil detection and classification is


49 available at https://huggingface.co/spaces/adamartukuli/RRWIT.


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.17.767224; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.17.767224)

(which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC
105 and is also made available for use under a CC0 license.

# 50 1 Introduction


51 Curculionidae (Coleoptera), the “true weevils”, is among the most diverse insect families,


52 with more than 6,000 genera and 97,000 described species &#91;1&#93;. Their diversification is closely


53 associated with flowering plants, and many species have a characteristic rostrum bearing


54 modified mouthparts used in feeding and oviposition &#91;1,2&#93;. This diversity includes major


55 agricultural pests, including the pecan weevil, _Curculio caryae_ (Horn), which is a serious pest of


56 pecan production and is monitored with Circle traps (Great Lake IPM, Vestaburg, MI) to guide


57 management &#91;3-7&#93;, and the boll weevil, _Anthonomus grandis grandis_ Boheman, a pest of


58 commercial cotton and is monitored with Pheromone Lure equipped traps (Technical Precision


59 Plastics, Mebane, NC) to guide management, remains subject to surveillance and eradication


60 activities in the Lower Rio Grande Valley of Texas along the U.S.-Mexico border &#91;8,9&#93;.


61 A persistent challenge in pecan and boll weevil surveillance is the frequent capture of


62 morphologically similar, non-pest weevil species in monitoring traps. Closely related taxa can


63 share body size, coloration, overall shape, and other external characters, while reliable species


64 identification may depend on subtle morphological differences requiring specialized taxonomic


65 expertise. Incorrect identification of a non-pest weevil as an economically important pest can


66 trigger unnecessary remedial actions, leading to excess pesticide use, disrupted trade, and


67 increased operational costs. Conversely, incorrect identification of an economic pest as a non

68 pest can lead to significant losses in yield or increased costs due to treatment delays.


69 Therefore, accurate, rapid, and field-deployable identification tools are critically needed to


70 improve management decisions and to reduce the costs associated with pest misclassification


71 while reserving uncertain cases for specialist examinations. Moreover, no existing AI models


72 have been specifically tailored to distinguish between pest and non-pest weevils that are


73 morphologically very similar and captured as by-catch in monitoring traps. Deep learning has


74 been applied to cotton pests, boll-weevil classification under limited-data conditions, and other


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.17.767224; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.17.767224)

(which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC
105 and is also made available for use under a CC0 license.


75 fine-grained insect-identification problems &#91;10-12&#93;. Image classification predicts an image-level


76 class, whereas object detection jointly predicts object identity and spatial extent, usually as a


77 bounding box &#91;13&#93;. For monitoring images that may contain small insects, overlapping targets,


78 debris, or heterogeneous backgrounds, localization and classification are therefore


79 complementary requirements &#91;14,15&#93;.


80 Agricultural detectors have addressed these challenges with feature fusion, attention, and


81 multi-scale processing, including AF-RCNN, AgriPest-YOLO, and Improved Pest-YOLO &#91;16-18&#93;.


82 Deployment-oriented studies further show that predictive accuracy must be considered together


83 with model size, computation, and measured hardware performance &#91;19,20&#93;. Model


84 interpretation adds another diagnostic dimension; attribution maps can indicate where model

85 associated evidence is concentrated, but they do not establish that a model has learned causal


86 biological characters. Recent lightweight agricultural studies increasingly combine comparative


87 benchmarking, ablation, computational characterization, and attribution visualization rather than


88 relying on a single accuracy value &#91;21&#93;.


89 These considerations motivated development of Rostrum Recognition: A Weevil


90 Identification Tool (RRWIT) by fine-tuning a nano-scale You Only Look Once version 11


91 detector (YOLO11n) &#91;22&#93; pretrained on Common Objects in Context (COCO) &#91;31&#93; for 14


92 morphologically similar pest and non-pest weevil taxa. Considering the dataset we have, we


93 hypothesized that transfer learning with a nano-scale one-stage detector would provide a useful


94 balance of fine-grained detection performance and computational efficiency for our finite


95 biological dataset. We tested that design by isolating the effect of pretrained initialization,


96 quantifying EigenGradCAM attribution localization, screening targeted architecture variants,


97 benchmarking eight representative detector frameworks under a common validation/test


98 contract, examining species-specific errors, and measuring model complexity, quantization


99 behavior, and controlled central processing unit (CPU) inference. The study was designed to


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.17.767224; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.17.767224)

(which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC
105 and is also made available for use under a CC0 license.


100 evaluate RRWIT as a reproducible and extensible framework for automated weevil


101 identification, not to treat the current web prototype as a validated final field-deployment system.


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.17.767224; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.17.767224)

(which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC
105 and is also made available for use under a CC0 license.

# 102 2 Materials and methods


103
## **2.1 Image acquisition and dataset composition**


104 Fourteen weevil taxa, including both pests and non-pests, were selected based on their


105 morphological similarity and frequency of capture in boll weevil and pecan weevil monitoring


106 traps. These included _A. grandis grandis_ (boll weevil) and eight non-pest weevil species, _A. g._


107 _thurberiae_ Pierce _, A. eugenii_ Cano, _A. texanus_ Dietz, _A. hunteri_ Burke and Cate, _A. fulvus_


108 LeConte, _A. cognatus_ Burke, _A. peninsularis_ Dietz, and _Conotrachelus seniculus_ LeConte.


109 Likewise, five key species included the pecan weevil ( _C. caryae_ ) set – the target pest species


110 and four non-pest weevils: _C. fulvus_ Chittenden _, C. proboscideus_ Fabricius _,_ _C. pardalis_


111 Chittenden, and _Conotrachelus naso_ LeConte.


112 Specimens were obtained with assistance from the Texas Boll Weevil Eradication


113 Foundation, Texas Department of Agriculture, Texas Pecan Growers Association, and cotton


114 and pecan producers across Texas. Additional contributions were made by the Texas A&amp;M


115 University Insect Collection. Images were collected under diverse conditions using tablets


116 (Samsung Galaxy S9 &#91;+&#93; ) and a camera mounted to a microscope (Olympus SZ40 with AmScope


117 camera). These devices were selected to simulate both field and laboratory environments,


118 thereby enhancing the generalizability of the dataset. Each specimen was photographed under


119 varying background and lighting conditions. Additional images were obtained from Weevils of


120 Texas, Weevils of North America, iNaturalist, and BugGuide and were taxonomically verified


121 before inclusion &#91;23-26&#93;. Image and specimen metadata were managed with Field Book &#91;27&#93; to


122 ensure meticulous tracking and maintain data integrity throughout the project by assigning each


123 weevil image a unique identifier. The curated collection contained more than 16,000 annotated


124 images distributed among the training, validation, and test partitions while keeping images from


125 the same biological specimen within unique subsets to avoid potential data leakage.


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.17.767224; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.17.767224)

(which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC
105 and is also made available for use under a CC0 license.


126
## **2.2 Image preprocessing, annotation, and dataset**


127
## **partitioning**


128 Images from heterogeneous source formats were standardized to JPEG with Pillow &#91;28&#93;.


129 Initial object annotations were generated with Grounding DINO, an open-set detector that links


130 visual and language representations &#91;29&#93;. Every generated bounding box was then inspected


131 and manually corrected in Roboflow annotation platform &#91;30&#93;. Corrections included box position


132 and dimensions, deletion of erroneous boxes, addition of missed objects, and verification of


133 taxonomic labels. The resulting expert-reviewed boxes and labels constituted the ground truth.


134 The annotated images were partitioned once before model development by taxon-stratified


135 sampling into training (80%), validation (10%), and test (10%) subsets (S1 Fig); partition


136 membership was retained for all subsequent analyses. Training images alone were used for


137 parameter estimation and stochastic augmentation. Augmentations were generated at training


138 during optimization rather than as pre-generated images, so augmented copies were never


139 distributed across partition subsets. Validation data were used for training monitoring, early


140 stopping, checkpoint selection, architecture screening, and, in the later framework benchmark,


141 model-specific confidence-threshold selection. The test partition was not used for fitting or


142 selection. Before the final benchmark accessed test annotations, all checkpoints, thresholds,


143 and evaluator logic were frozen from validation data, and the evaluator was required to


144 reproduce validation metrics while technically unable to open the with-held test annotation


145 subset.


146
## **2.3 Development of RRWIT**


147 RRWIT was developed by fine-tuning YOLO11n &#91;22&#93; from the Ultralytics yolo11n.pt


148 checkpoint pretrained on the COCO object-detection dataset &#91;31&#93;. The original run used random


149 seed 0 with deterministic training enabled. The archived configuration specified up to 1,000


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.17.767224; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.17.767224)

(which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC
105 and is also made available for use under a CC0 license.


150 epochs, early-stopping patience of 15 epochs, batch size 16, 640 × 640-pixel input, six data

151 loader workers, caching, the complete training partition, and automatic mixed-precision training.


152 Optimizer selection was performed using the Ultralytics automatic optimizer routine


153 intentionally to ensure reproducibility and consistency. The archived training configuration


154 specified an initial learning rate (lr0) of 0.01, a final learning-rate factor (lrf) of 0.01, momentum


155 of 0.937, weight decay of 0.0005, three warm-up epochs, an initial warm-up momentum of 0.8,


156 and an initial warm-up bias learning rate of 0.1. Because the Ultralytics automatic optimizer


157 routine can internally select the optimizer and override some user-specified optimization


158 parameters, these values are reported as the archived configuration settings rather than as the


159 effective optimizer settings used during training. Loss weights were set to 7.5 for box


160 regression, 0.5 for classification, and 1.5 for distribution focal loss. No layers were frozen;


161 dropout was 0; rectangular and multi-scale training were disabled.

### 162 2.3.1 Training-time augmentation


163 Stochastic augmentation was restricted to the training partition. The seed-0 configuration


164 used hue = 0.015, saturation = 0.7, brightness/value = 0.4, translation = 0.1, scale = 0.5,


165 horizontal-flip probability = 0.5, and Mosaic probability = 1.0, with Mosaic disabled during the


166 final 10 epochs. Rotation, shear, perspective transformation, vertical flipping, MixUp, copy

167 paste, BGR channel swapping, and multi-scale training were disabled. Validation and test


168 images were evaluated without stochastic training augmentation.


169 Training used an NVIDIA A100-SXM4-80GB graphics processing unit (GPU) on the USDA


170 Agricultural Research Service SCINet high-performance-computing resource. The selected


171 checkpoint was frozen as RRWIT.v1.0.0.pt (referred as RRWIT in this manuscript) and served


172 as the seed-0 model for diagnostic analyses and prototype deployment. A demonstration web


173 application interface was implemented with Gradio and hosted on Hugging Face Spaces


174 &#91;32,33&#93;.


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.17.767224; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.17.767224)

(which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC
105 and is also made available for use under a CC0 license.


175

## 176 2.4 Random-initialization and variance-matched transfer

177
## **learning experiment**


178 To isolate pretrained initialization from detector architecture, YOLO11n models initialized


179 from random weights were compared with models initialized from the same frozen COCO

180 pretrained yolo11n.pt source. Both conditions used the fixed dataset partitions and prespecified


181 seeds 42, 123, and 2025. Random-initialization models were instantiated directly from


182 yolo11n.yaml without loading pretrained weights. Transfer-learning replicates were initialized


183 independently from yolo11n.pt rather than from the seed-0 RRWIT checkpoint. The principal


184 RRWIT training contract was retained, including the 1,000-epoch maximum, 640 × 640 input,


185 batch size 16, patience 15, deterministic execution, automatic mixed precision and optimizer


186 selection, weight decay, warm-up, loss weights, Mosaic schedule, horizontal flipping, and


187 disabled MixUp. Each run was an independent training replicate; validation selected the


188 checkpoint, and test evaluation occurred only afterward. Results were summarized as arithmetic


189 mean ± sample standard deviation (SD) across seeds.


190
## **2.5 Quantitative EigenGradCAM analysis**


191 The frozen RRWIT checkpoint was examined using EigenGradCAM implemented in


192 pytorch-grad-cam &#91;34&#93;. The purpose of this explainable artificial intelligence (XAI) analysis was


193 to quantify attribution localization relative to the annotated insect and determine whether


194 attribution localization varied with object size or prediction outcome. Inference used an input


195 resolution of 640 × 640 pixels, a confidence threshold of 0.25, and a non-maximum-suppression


196 intersection over union (IoU) threshold of 0.70. For this audit, a prediction was considered


197 correct only when the predicted species was correct and the detection matched the relevant


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.17.767224; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.17.767224)

(which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC
105 and is also made available for use under a CC0 license.


198 ground-truth bounding box at IoU ≥0.50. Class-targeted EigenGradCAM attribution was


199 calculated from YOLO11n model layer 16, corresponding to the high-resolution P3-scale feature


200 pathway &#91;34–37&#93;.


201 The test partition contained 1,693 images, of which 1,685 contained a single annotated


202 object and formed the audit population. Each image was classified as a correct prediction,


203 species misclassification, localization failure, or no detection under the frozen evaluation rules.


204 Ground-truth object size was calculated as bounding-box area divided by image area and


205 categorized as tiny (&lt;1%), small (1–&lt;10%), medium (10–&lt;50%), or large (≥50%). The complete


206 1,685-image audit was used to characterize prediction outcomes independently of the attribution


207 subset. Because EigenGradCAM required an available detection, no-detection cases remained


208 in the performance audit but were excluded from attribution calculations.


209 To preserve error diversity while limiting dominance by the more abundant correct


210 predictions, all qualifying detected incorrect cases were retained, whereas correctly detected


211 cases were ranked by prediction confidence within each species-by-size stratum and limited to


212 a maximum of three observations per stratum. This deterministic procedure produced a fixed


213 attribution subset of 258 images, comprising 115 correct and 143 incorrect detections. The


214 deliberate enrichment of incorrect predictions was used to compare attribution behavior


215 between successful and unsuccessful predictions and was not used to estimate the test-set


216 error rate. Attribution maps were normalized within each image before quantitative analysis.


217 Ground-truth-box enrichment was calculated as the fraction of total EigenGradCAM


218 attribution within the annotated bounding box divided by the fraction of image area occupied by


219 that box. An enrichment value of 1 indicates attribution proportional to box area, values &gt;1


220 indicate relative concentration of attribution within the insect box, and values &lt;1 indicate relative


221 under-enrichment within the box. Because bounding-box area contributes directly to the


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.17.767224; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.17.767224)

(which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC
105 and is also made available for use under a CC0 license.


222 normalization, enrichment was interpreted jointly with continuous object-area fraction.


223 Representative examples were selected without visual-preference screening: among correctly


224 detected in-house images, the observation closest to the median enrichment within each size


225 stratum was selected, with IoU, prediction confidence, and filename used sequentially as


226 deterministic tie-breakers. Associations between continuous object-area fraction and enrichment


227 were assessed separately for correct and incorrect predictions using Spearman rank correlation,


228 with 5,000 bootstrap replicates used where uncertainty intervals were reported. Attribution maps


229 were interpreted as diagnostic evidence of model attention patterns rather than as evidence of


230 causal taxonomic characters &#91;35–37&#93;.


231
## **2.6 Targeted architecture ablation**


232 EigenGradCAM motivated a controlled test of whether additional shallow-scale information


233 or explicit channel-spatial recalibration improved the RRWIT YOLO11n baseline. For this initial


234 architecture screen, all four models were trained using the same random seed 42 to hold the


235 seed constant while comparing architectural modifications. The four variants were A0, the


236 unmodified COCO-pretrained baseline; A1, a P2-to-P3 feature-assistance pathway; A2, a


237 Convolutional Block Attention Module (CBAM) inserted at P3; and A3, the combined P2

238 assistance plus P3-CBAM design. The CBAM module sequentially applies channel and spatial


239 attention &#91;38&#93;. All variants retained P3, P4, and P5 detection outputs and were trained using the


240 same fixed training and validation partitions and RRWIT protocol. Selection used mean average


241 precision across intersection-over-union (IoU) thresholds from 0.50 to 0.95 (mAP50-95) as the


242 primary criterion, followed by mean average precision at an IoU threshold of 0.50 (mAP50) and


243 fewer parameters as tie-breakers. The test split was not accessed during screening. Parameter


244 counts were obtained from the exact model implementations, and theoretical computational


245 complexity, expressed as billions of floating-point operations (GFLOPs) at an input resolution of


246 640 × 640 pixels, was estimated using the Ultralytics get_flops implementation. A modified


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.17.767224; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.17.767224)

(which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC
105 and is also made available for use under a CC0 license.


247 architecture would advance to multi-seed evaluation and subsequent test-set confirmation only


248 if it surpassed A0 on the prespecified validation criteria.


249
## **2.7 Eight-framework, three-seed benchmark**


250 RRWIT was compared with seven representative detector configurations: YOLOv5nu,


251 YOLOv8n, YOLO12n, YOLO26n, Single Shot MultiBox Detector Lite (SSDLite320)


252 MobileNetV3-Large, EfficientDet-D0, and Faster R-CNN MobileNetV3-Large-320 feature


253 pyramid network (FPN) &#91;39–46&#93;. Published architecture and implementation characteristics are


254 summarized in S1 Table. The benchmark was designed to compare complete framework

255 specific detector configurations on the same biological task rather than to isolate architecture


256 under an identical input-resolution and optimization regime. Accordingly, reference or


257 architecture-appropriate input resolutions and family-specific training configurations were


258 retained rather than forcing all detector families to use a single resolution, optimizer, or batch


259 size. The YOLO models used 640 × 640-pixel inputs and batch size 16, SSDLite320 used 320 ×


260 320 pixels and batch size 32, EfficientDet-D0 used 512 × 512 pixels and batch size 16, and


261 Faster R-CNN MobileNetV3-Large-320 FPN used its low-resolution 320-pixel configuration with


262 batch size 8. All frameworks used the same frozen RRWIT source-image partitions and


263 common evaluation definitions. Because input resolution and training configuration can


264 influence both predictive performance and computational requirements, results were interpreted


265 as comparisons among the evaluated framework-specific configurations rather than as an iso

266 resolution architecture ablation.


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.17.767224; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.17.767224)

(which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC
105 and is also made available for use under a CC0 license.


267 Each framework was trained independently with seeds 42, 123, and 2025, producing 24


268 model-seed combinations on identical source-image partitions and 14 classes. Framework


269 adapters transformed the same annotations into the coordinate and label conventions required


270 by each implementation, and reconstruction audits checked box equivalence. Validation and


271 test preprocessing were deterministic. For each model-seed pair, canonical checkpoint selection


272 used validation COCO-style mAP50-95 as the primary criterion &#91;31&#93;; confidence thresholds


273 required for F1 metrics were also selected using validation predictions. Before test access, a


274 locked gate reloaded every selected checkpoint through its framework-specific inference path


275 and reproduced canonical validation metrics and locked-threshold F1 values within prespecified


276 tolerances. The test-enabled evaluator was authorized only after all 24 gates passed. Test


277 results were not subsequently used to retrain models, substitute checkpoints, adjust thresholds,


278 or select architectures.


279


280
## **2.8 Performance metrics and statistical analysis**


281 The primary benchmark metric was COCO-style mean average precision averaged over


282 IoU thresholds 0.50 to 0.95 in increments of 0.05 (mAP50-95) &#91;31&#93;. Additional metrics included


283 mAP at IoU 0.50 (mAP50), mAP at IoU 0.75 (mAP75), and average recall with at most 100


284 detections per image (AR100). At validation-selected confidence thresholds, precision and recall


285 were used to derive macro, micro, and support-weighted F1 scores; species-specific AP50-95,


286 AP50, and F1 were retained to prevent aggregate metrics from obscuring low-support taxa.


287 Three-seed experiments were summarized as arithmetic mean ± sample standard deviation


288 (SD; n - 1 degrees of freedom). Because each detector had only three independent training


289 replicates, small between-model differences were interpreted descriptively rather than as strong


290 inferential evidence; no hypothesis test was used to promote a model based on locked-test


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.17.767224; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.17.767224)

(which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC
105 and is also made available for use under a CC0 license.


291 results. EigenGradCAM area-enrichment associations used Spearman rank correlation and,


292 where reported, 5,000 bootstrap replicates.


293
## **2.9 Computational footprint analysis**


294 Computational characteristics were measured from the exact trained 14-class models.


295 Registered model parameters were counted directly, and 32-bit floating-point (FP32) parameter


296 memory was calculated from registered parameter tensors stored in 32-bit floating point,


297 excluding buffers, optimizer/scheduler state, training history, metadata, and serialization


298 overhead. Theoretical computation was estimated for a batch-size-one eager forward pass at


299 each framework’s frozen input resolution with torch.utils.flop_counter.FlopCounterMode &#91;34&#93; and


300 reported as GFLOPs. GFLOPs were treated as an operation-count descriptor, not as latency,


301 throughput, energy use, or hardware-specific speed. Raw framework-native training checkpoint


302 sizes were not compared because their serialized contents were not equivalent across families;


303 deployment artifact size was evaluated separately after export to a common representation.


304
## **2.10 Deployment-oriented ONNX and INT8 evaluation**


305 Four representative lightweight detectors were carried forward after the accuracy


306 benchmark: YOLO26n seed 123; YOLO12n seed 42; RRWIT seed 2025; and EfficientDet-D0


307 seed 42. Each checkpoint was selected using validation performance only and was analyzed


308 separately from the three-seed benchmark means. Models were exported in full precision to


309 Open Neural Network Exchange (ONNX) opset 17 &#91;47&#93; and audited for structural validity and


310 output parity. Static post-training 8-bit integer (INT8) quantization used ONNX Runtime &#91;48&#93; with


311 QuantizeLinear/DeQuantizeLinear representation, signed 8-bit activations and weights (S8S8),


312 per-tensor weight quantization, and MinMax calibration. Calibration used a frozen cohort of 512


313 training images; validation and test images were excluded.


314 Common-protocol INT8 artifacts were evaluated on validation data for mAP50-95, mAP50,


315 support-weighted F1, proportional mAP50-95 retention, serialized artifact size, and percentage


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.17.767224; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.17.767224)

(which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC
105 and is also made available for use under a CC0 license.


316 size reduction. A prespecified aggregate accuracy gate required absolute declines ≤0.010 in


317 mAP50-95, mAP50, and weighted F1. Because EfficientDet-D0 degraded substantially under


318 the common S8S8/per-tensor protocol, additional signed/unsigned activation representations


319 and per-channel weight quantization were evaluated as architecture-specific diagnostic


320 sensitivity analyses only; they were not substituted for the authoritative common-protocol result.

### 321 2.10.1 Controlled CPU inference


322 FP32 and common-protocol INT8 artifacts were benchmarked with ONNX Runtime


323 CPUExecutionProvider. The batch-size-one protocol used four intra-operation threads, 16


324 warm-up images, and three timed repetitions of the same 128 images drawn exclusively from


325 the training partition. Model-core latency measured ONNX graph execution after input


326 preparation; end-to-end latency additionally included image loading/decoding, family-canonical


327 preprocessing, ONNX inference, detector-specific decoding and non-maximum suppression


328 where required, and conversion to the common prediction representation. Median latency, 95th

329 percentile end-to-end latency, throughput, and peak process memory were retained. Results


330 characterize this x86 CPU/ONNX Runtime configuration and were not extrapolated to ARM


331 processors, neural-processing units, GPUs, or other edge accelerators.


332
## **2.11 Software environment reproducibility and code**


333
## **availability.**


334 The RRWIT model used Python 3.11.7, PyTorch 2.5.1 with CUDA 12.4, and Ultralytics


335 8.3.85. Architecture screening used Python 3.10.20, PyTorch 2.5.1 with CUDA 12.4, and


336 Ultralytics 8.3.28. The frozen cross-framework benchmark used Python 3.10.20, PyTorch 2.7.1


337 with CUDA 12.6, Torchvision 0.22.1, Ultralytics 8.4.118, effdet 0.4.1, timm 0.9.16, pycocotools


338 2.0.11, ONNX 1.22.0, and ONNX Runtime 1.23.2 to accommodate framework-specific


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.17.767224; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.17.767224)

(which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC
105 and is also made available for use under a CC0 license.


339 dependency constraints. Requirement files and codes used in the main analysis are available


340 at &#91;https://github.com/adama960/A-Weevil-Identification-Tool-RRWIT-&#93;(https://github.com/adama960/A-Weevil-Identification-Tool-RRWIT-) GitHub repository.


341 The final fine-tuned YOLOv11n-based model checkpoint is publicly available on our Hugging


342 Face Space.

# 343 3 Results and Discussion


344
## **3.1 RRWIT achieves robust overall species-level**


345
## **detection across morphologically similar weevils**


346 Manual annotation review and dataset-level diagnostics showed representation of all 14


347 taxa together with broad variation in normalized object position and bounding-box dimensions


348 (S1 and S2 Figs). Class support remained uneven: _A. grandis grandis_ was the largest class,


349 whereas several taxa had substantially fewer images; most notably _C. fulvus_ (S1 Fig).


350 Bounding-box centers were predominantly specimen-centered but occupied a wider region of


351 the image plane, and width-height distributions covered a broad range. The expected positive


352 width-height association and geometric constraints between box size and valid center locations


353 were consistent with normalized object annotations (S2 Fig). The seed-0 RRWIT training run


354 showed progressive convergence before early stopping at epoch 565 (Fig 1). Training and


355 validation box, classification, and distribution focal losses generally decreased over training (Fig


356 1A and 1B), while validation mAP50 and mAP50-95 increased and gradually approached a


357 plateau (Fig 1C). The checkpoint at epoch 550 was selected on the basis of validation


358 performance. Although training losses continued to decline, the validation losses did not show


359 pronounced late-stage deterioration, and validation mean average precision remained near its


360 highest levels toward the end of training. Thus, the learning curves did not indicate substantial


361 late-stage overfitting in this run.


362


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.17.767224; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.17.767224)

(which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC
105 and is also made available for use under a CC0 license.


363 **Fig 1. Training convergence and validation performance of Rostrum Recognition:**


364 **A Weevil Identification Tool (RRWIT).** (A) Training bounding-box regression loss (box_loss),


365 classification loss (cls_loss), and distribution focal loss (dfl_loss). (B) Corresponding validation


366 losses. (C) Validation mean average precision at an intersection-over-union (IoU) threshold of


367 0.50 (mAP50) and mean average precision averaged across IoU thresholds from 0.50 to 0.95


368 (mAP50-95). The vertical dashed line marks the validation-selected checkpoint at epoch 550;


369 training terminated at epoch 565 under the early-stopping criterion.


370


371 Evaluation of the validation and held-out test partitions showed strong overall detection


372 performance, with some reduction in performance on the previously unseen test data (Fig 2).


373 Overall mAP50 was 0.96 on the validation partition and 0.87 on the test partition. The precision

374 recall curves likewise showed generally strong performance across taxa, although the degree of


375 precision-recall trade-off varied among species (Fig 2C and 2D). The all-class F1 curve reached


376 a maximum of approximately 0.84 at a confidence threshold of 0.464 on validation data and


377 0.81 at 0.393 on test data (Fig 2E and 2F). Species-level precision, recall, and F1 curves


378 showed that performance was not uniform across taxa, with several classes maintaining


379 comparatively high values across a broad range of confidence thresholds while others showed


380 greater sensitivity to threshold selection (Fig 2A, 2B, and 2E–H). Together, the validation-to-test


381 changes indicate that RRWIT retained substantial discriminatory performance on held-out


382 previously unseen test partitions subset data while also revealing species-specific differences


383 that are obscured by aggregate mAP values.


384 **Fig 2. Precision, recall, F1, and precision-recall performance of RRWIT on the**


385 **validation and held-out test partitions subset data.** Left-column panels show validation


386 performance and right-column panels show held-out previously unseen test performance. (A, B)


387 Precision as a function of confidence threshold; (C, D) precision-recall curves; (E, F) F1 score


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.17.767224; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.17.767224)

(which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC
105 and is also made available for use under a CC0 license.


388 as a function of confidence threshold; and (G, H) recall as a function of confidence threshold.


389 Thin colored lines represent individual weevil taxa, and the thick blue line represents the all

390 class summary. Values reported with the all-class F1 curves indicate the maximum F1 score


391 and corresponding confidence threshold. Values reported in the precision-recall legends are


392 average precision at an intersection-over-union threshold of 0.50 (AP50), with the thick blue


393 curve reporting mean average precision across classes (mAP50).


394 The normalized confusion matrices further resolved species-specific differences in


395 RRWIT performance (Fig 3). Correct classifications dominated the diagonal for most taxa in


396 both validation and test data, but the magnitude of the diagonal varied among species. On the


397 validation partition, diagonal proportions ranged from 0.54 to 1.00, with particularly high values


398 for _Curculio fulvus_ (1.00), _Anthonomus grandis grandis_ (0.98), _Anthonomus texanus_ (0.95), and


399 _Anthonomus fulvus_ (0.94) (Fig 3A). On the independent held-out previously unseen test


400 partition, diagonal proportions ranged from 0.50 to 0.98, with high values for _Anthonomus_


401 _grandis thurberiae_ (0.98), _Anthonomus grandis grandis_ (0.97), _Anthonomus texanus_ (0.94), and


402 _Anthonomus fulvus_ (0.92) (Fig 3B). Lower diagonal proportions were observed for some taxa,


403 including _Curculio fulvus_ (0.50), _Curculio caryae_ (0.59), _Anthonomus cognatus_ (0.61),


404 _Anthonomus hunteri_ (0.62), and _Conotrachelus seniculus_ (0.62) on the test partition. The off

405 diagonal structure shows that errors were concentrated among particular taxon pairs rather than


406 being uniformly distributed across the 14 classes. These class-specific patterns complement the


407 aggregate mAP and F1 results by identifying taxa for which discrimination remained more


408 challenging. This pattern is consistent with the subtle characters used in weevil taxonomy


409 &#91;49,50&#93;, although class representation, orientation, illumination, scale, and background can also


410 contribute to errors &#91;16-18&#93;.


411 **Fig 3. Normalized confusion matrices for RRWIT on the validation and held-out**


412 **previously unseen test partitions subset data.** (A) Validation partition. (B) Independent test


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.17.767224; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.17.767224)

(which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC
105 and is also made available for use under a CC0 license.


413 partition. Rows represent predicted classes and columns represent true classes. Values are


414 normalized within each true-class column and therefore represent the proportion of instances for


415 each true class assigned to each predicted class under the detection-matching criteria. Diagonal


416 cells indicate correct class assignments, whereas off-diagonal cells indicate misclassifications.


417


418 To further evaluate the robustness of the RRWIT model, a qualitative visual analysis was


419 conducted on representative individual and batches of images (Fig 4 and S3 Fig). These results


420 highlight both the localization and classification performance of the model across several


421 morphologically similar species. As shown in Fig 4, ground truth annotations (Fig 4A) reflect the


422 diversity in image acquisition. The corresponding model predictions (Fig 4B) demonstrate that


423 the model reliably delineated bounding boxes and correctly assigned species labels with high


424 confidence across most samples. Colors correspond to species classes, with label confidence


425 shown numerically in the predictions. In the representative batch shown in Fig 4, 15 of 16


426 images were correctly classified, with one misclassification involving _A. eugenii_ (green yellow


427 color), underscoring the model's capability when species-specific features are distinct and


428 imaging conditions are optimized. When RRWIT is tested with a single image (S3 Fig), it


429 automatically detects the presence of weevils, generates bounding boxes, and assigns a


430 species identification based on the trained detection model. The predicted species label and


431 associated confidence score are displayed directly on the image, allowing the user to visually


432 confirm the result. Overall, the qualitative inspection is consistent with the quantitative


433 evaluations in Figs 1–4 while also illustrating model errors under challenging image conditions.


434 RRWIT demonstrates detection performance under various image conditions, including lighting


435 variation and occlusion. Notable misclassifications are evident in some panels due to species


436 similarity and challenging visual contexts. These results highlight both model strengths and


437 limitations under real-world conditions. It also emphasizes the importance of expanding image


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.17.767224; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.17.767224)

(which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC
105 and is also made available for use under a CC0 license.


438 diversity and representation, particularly for morphologically ambiguous or underrepresented


439 species.


440


441 **Fig 4. Ground truth vs. RRWIT predictions for one representative batch consisting**


442 **of 16 images.** (A) Displays ground truth labels, while (B) displays predicted bounding boxes


443 and class labels by the RRWIT. Colors correspond to species classes, with label confidence


444 shown numerically in the predictions.


445


446 In the broader agricultural context, reported accuracies vary substantially with taxonomic


447 granularity, object scale, annotation policy, and evaluation definition; direct cross-paper


448 numerical ranking can therefore be misleading &#91;51-55&#93;. RRWIT’s main contribution is the


449 application of lightweight object detection to a fine-grained pest/non-pest weevil identification


450 problem, with the subsequent controlled experiments providing the more rigorous basis for


451 comparative conclusions.


452 Cross-study comparisons provide context but require caution because agricultural detection


453 papers differ in taxonomic granularity, object scale, image acquisition, class balance, dataset


454 partitioning, and evaluation definitions. AF-RCNN reported 56.4% mAP on a 24-class


455 agricultural pest dataset &#91;16&#93;, AgriPest-YOLO reported 71.3% mAP on an approximately 25,000

456 image light-trap dataset &#91;17&#93;, and Improved Pest-YOLO reported 73.4% mAP after adding


457 efficient channel attention and transformer-based feature processing &#91;18&#93;. These studies


458 emphasize the difficulty of multi-species pest detection when targets are small, densely


459 distributed, or observed against complex backgrounds; their numerical values should not be


460 treated as direct rankings against RRWIT because the underlying tasks and metrics differ.


461 Recent work increasingly evaluates accuracy together with computational efficiency.


462 Faster-PestNet was developed as a lightweight crop-pest detector &#91;51&#93;, whereas PestLite


463 reported 90.7% mAP50 while reducing a YOLOv5-based model from 7.03 to 6.09 million


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.17.767224; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.17.767224)

(which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC
105 and is also made available for use under a CC0 license.


464 parameters &#91;52&#93;. A UAV-oriented pest detector evaluated on five broad pest groups reported


465 mean precision of 96.0%, recall of 93.0%, and mAP of 95.0% &#91;53&#93;. Those values exceed the


466 RRWIT test mAP50, but the UAV study separated five relatively broad pest groups, whereas


467 RRWIT distinguishes 14 morphologically similar weevil taxa, including congeneric pest and non

468 pest lookalikes. Differences in reported accuracy therefore reflect both detector performance


469 and task difficulty.


470 The distinction between mAP50 and stricter localization metrics is also increasingly explicit


471 in contemporary agricultural vision. A rice-pest study reported 98.8% mAP50 and 78.6%


472 mAP50-95 &#91;54&#93;. YOLO-DCPG, developed for intensive small-target pest detection on Pest24,


473 reported 74.0% mAP50 and 47.5% mAP50-95 with 1.47 million parameters and 5.7 GFLOPs


474 &#91;55&#93;. Together, these studies support reporting multiple complementary metrics rather than


475 presenting a single mAP value as a complete description of detector behavior. RRWIT


476 contributes to this literature from a distinct fine-grained taxonomic perspective, where


477 operational value depends on separating pest species from visually similar by-catch rather than


478 only detecting visually dissimilar pest categories.


479 A trial RRWIT web application was implemented as a user-facing web interface at


480 https://huggingface.co/spaces/adamartukuli/RRWIT. Users can upload an image or use live


481 image capture and receive predicted bounding boxes, species labels, confidence values, and an


482 adjustable confidence threshold (S1 Video). The interface demonstrates how the trained


483 detector can be presented for interactive expert-assisted identification. It is a functional


484 prototype, not evidence of independent field validation. The present study did not systematically


485 measure application-level latency, user performance, scalability, energy consumption, on

486 device, performance on representative mobile and edge hardware.


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.17.767224; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.17.767224)

(which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC
105 and is also made available for use under a CC0 license.


487
## **3.2 Transfer learning improves detection performance**


488
## **and training consistency**


489 Randomly initialized YOLO11n models learned the task, with declining training and


490 validation losses and increasing validation mAP across all three seeds, although their training


491 trajectories and validation-selected stopping points varied among runs (Fig 5). These results


492 show that random initialization did not prevent optimization of the YOLO11n architecture and


493 provide a matched baseline for evaluating the contribution of pretrained initialization to held-out


494 performance and between-seed variability.


495 **Fig 5. Three-seed training convergence of randomly initialized YOLO11n controls.**


496 Seeds 42, 123, and 2025 were trained under the matched RRWIT training protocol. (A) Training


497 box, classification, and distribution focal losses. (B) Corresponding validation losses. (C)


498 Validation mAP50 and mAP50-95. Vertical lines mark the validation-selected checkpoint for


499 each seed.


500 On the held-out test partition, the randomly initialized models achieved mAP50 values of


501 0.798, 0.723, and 0.762 for seeds 42, 123, and 2025, respectively, corresponding to a mean ±


502 sample standard deviation (SD) of 0.761 ± 0.038 (Fig 6A). Their corresponding mAP50-95


503 values were 0.616, 0.550, and 0.590, with a mean of 0.586 ± 0.033. Precision, recall, and F1


504 also varied among seeds (Fig 6A), while species-level AP50 and AP50-95 showed substantial


505 heterogeneity in both average performance and between-seed variability (Fig 6B). Together,


506 these results demonstrate that the randomly initialized models could learn the detection task but


507 exhibited lower performance and noticeable variation in held-out performance across the three


508 seeds.


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.17.767224; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.17.767224)

(which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC
105 and is also made available for use under a CC0 license.


509 **Fig 6. Held-out test performance of three randomly initialized YOLO11n controls.**


510 (A) Overall precision, recall, F1, mAP50, and mAP50-95 for seeds 42, 123, and 2025; open


511 symbols and error bars indicate the arithmetic mean ± sample SD across the three seeds. (B)


512 Species-level AP50 and AP50-95 summarized as mean ± sample SD across the same three


513 seeds.


514 The matched COCO-pretrained YOLO11n models achieved test mAP50 values of 0.803,


515 0.836, and 0.799 for seeds 42, 123, and 2025, respectively, corresponding to 0.813 ± 0.020,


516 and mAP50-95 values of 0.622, 0.645, and 0.617, corresponding to 0.628 ± 0.015 (Fig 7A).


517 Relative to random initialization, pretrained initialization therefore increased mean mAP50 by


518 0.052 and mean mAP50-95 by 0.042. Between-seed variability was also lower for the pretrained


519 models; the sample SD decreased from 0.038 to 0.020 for mAP50 and from 0.033 to 0.015 for


520 mAP50-95. The mAP50-95 difference favored pretrained initialization for each matched seed,


521 with the largest difference observed for seed 123, which had the lowest randomly initialized


522 performance. Given the small number of replicated seeds (n = 3 per initialization condition),


523 these differences are interpreted descriptively rather than as evidence of statistical significance.


524 Species-level AP remained heterogeneous after pretrained initialization (Fig 7B), indicating that


525 the overall improvement did not eliminate differences in detection difficulty among taxa. The


526 three replicated pretrained models did not reproduce the performance of the seed-0 RRWIT


527 checkpoint, illustrating the stochastic variability associated with model training. The seed-0


528 checkpoint was retained as the deployed RRWIT model, whereas the subsequent three-seed


529 experiment was designed as an independent, matched comparison of pretrained and randomly


530 initialized models and therefore was interpreted on the basis of paired initialization conditions


531 rather than comparison with seed-0 RRWIT.


532 **Fig 7. Held-out test performance of three variance-matched COCO-pretrained**


533 **YOLO11n models.** (A) Overall precision, recall, F1, mAP50, and mAP50-95 for seeds 42, 123,


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.17.767224; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.17.767224)

(which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC
105 and is also made available for use under a CC0 license.


534 and 2025; open symbols and error bars indicate the arithmetic mean ± sample SD across the


535 three seeds. (B) Species-level AP50 and AP50-95 summarized as mean ± sample SD across


536 the same three seeds; labels report the number of test instances (n) for each taxon.


537 The direction of these matched results is consistent with agricultural transfer-learning


538 studies showing benefits of pretrained representations when target-domain training data are


539 limited &#91;56,57&#93; and with object-detection evidence that appropriate reuse of pretrained visual


540 features can improve downstream detection &#91;58&#93;. For RRWIT, the controlled comparison


541 indicates that COCO-pretrained initialization contributed to both higher mean held-out detection


542 performance and lower between-seed variability relative to random initialization under the


543 matched training protocol. The persistence of species-level performance differences, however,


544 indicates that pretrained initialization did not eliminate taxon-specific detection challenges.


545
## **3.3 EigenGradCAM reveals object-size-dependent** 546 attribution localization and classification-dominated


547
## **failure patterns**


548 The complete 1,685-image single-object audit separated species assignment from spatial


549 localization. RRWIT correctly classified and localized 1,527 images (90.6%); among the


550 remaining images, 137 were species misclassifications, six were localization failures, and 15


551 had no detection. Thus, residual errors were dominated by fine-grained species assignment


552 rather than inability to place a bounding box around the insect. The proportion correctly


553 classified and localized was 90.4% for tiny objects, 93.6% for small, 93.4% for medium, and


554 83.6% for large objects. These RRWIT-specific size strata were defined by relative ground-truth


555 box area (&lt;1%, 1–&lt;10%, 10–&lt;50%, and ≥50% of image area, respectively), and the metric


556 required both correct species assignment and localization. The lower percentage in the large

&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.17.767224; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.17.767224)

(which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC
105 and is also made available for use under a CC0 license.


557 object stratum therefore should not be interpreted as evidence that large objects were


558 intrinsically more difficult to detect; rather, prediction success in this seed-0 RRWIT audit did not


559 vary monotonically with relative object area. EigenGradCAM attribution was evaluated on the


560 fixed error-enriched subset of 258 detected images and summarized in Fig 8. Representative


561 maps were selected objectively by median enrichment within each size stratum. Ground-truth

562 box enrichment was highest and most variable for tiny objects and approached the area

563 proportional reference of 1 as object area increased. Continuous object area was strongly and


564 inversely associated with enrichment for both correct predictions (Spearman rho = -0.71, n =


565 115) and incorrect predictions (rho = -0.80, n = 143). Because enrichment divides attribution


566 fraction by box-area fraction, large values for tiny objects should not be interpreted as stronger


567 or more biologically meaningful attention.


568 **Fig 8. EigenGradCAM analysis of object-size-dependent attribution localization in**


569 **RRWIT.** (A-D) Objectively selected correctly detected examples for tiny, small, medium, and


570 large ground-truth objects, showing the original image, prediction/ground truth, and class

571 targeted EigenGradCAM map from layer 16. (E) Ground-truth-box enrichment distributions for


572 correct and incorrect detections in the 258-image error-enriched subset. (F) Continuous object


573 area versus enrichment. Enrichment decreased with object area for correct (Spearman rho = 

574 0.71, n = 115) and incorrect detections (rho = -0.80, n = 143). Attribution intensity is normalized


575 within each image and is not an absolute cross-image scale.


576


577 Correct and incorrect predictions overlapped substantially in their localization patterns. An


578 attribution map can therefore be concentrated on the insect even when the species label is


579 wrong. This is expected for a fine-grained task in which a detector may identify the biologically


580 relevant object yet fail to weight the specific visual characters needed for taxon discrimination.


581 Grad-CAM-type methods are post hoc attribution tools, not causal explanations, and visual


582 saliency can be misleading without quantitative controls &#91;35,37&#93;. The present analysis therefore


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.17.767224; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.17.767224)

(which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC
105 and is also made available for use under a CC0 license.


583 supports a limited conclusion; RRWIT frequently derives class-associated evidence from the


584 specimen region, but attribution localization alone does not establish taxonomically correct or


585 biologically causal feature use.


586 The predominance of species-assignment errors explains why correct and incorrect


587 detections can show similar object-focused attribution. The full audit contained recurrent


588 substitutions such as _Conotrachelus seniculus_ to _C. naso_ and _Anthonomus peninsularis_ to _A._


589 _texanus_, and the focal _Curculio_ audit showed repeated confusion with _C. proboscideus_ . These


590 observations are consistent with the class-confusion structure in Fig 3 and with taxonomic


591 literature describing subtle morphological distinctions within the _Anthonomus_ and _Curculio_


592 genera &#91;49,50&#93;. Collectively, these observations support diagnostic interpretation; RRWIT often


593 localizes the relevant specimen correctly, but some residual failures occur at the finer species

594 discrimination stage rather than at gross object localization.


595
## **3.4 Targeted feature-scale and attention**


596
## **modifications do not outperform baseline RRWIT**


597 Multi-scale feature fusion and attention are frequently used in agricultural detection when


598 targets vary strongly in size or weak visual features must be amplified. Recent literature studies


599 have likewise treated feature-fusion and attention modules as testable architectural hypotheses


600 and evaluated them by ablation against an unchanged baseline &#91;59,60&#93;. We therefore used the


601 object-size-dependent EigenGradCAM pattern to initiate, rather than predetermine, the RRWIT


602 architecture screen.


603 The attribution analysis initiated a direct architectural hypothesis: additional shallow-scale


604 information or explicit channel-spatial recalibration might improve performance if spatial feature


605 preservation were a principal limitation. However, the seed-42 validation screen did not support


606 that hypothesis. Baseline A0 achieved precision 0.801, recall 0.791, mAP50 0.842, and mAP50

607 95 0.648. A1 (P2-to-P3 assistance) fell to 0.672, 0.656, 0.718, and 0.536; A2 (P3 CBAM)


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.17.767224; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.17.767224)

(which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC
105 and is also made available for use under a CC0 license.


608 reached 0.786, 0.762, 0.820, and 0.635; and A3 (combined) reached 0.814, 0.685, 0.787, and


609 0.611. A0 therefore remained best on the prespecified primary metric (S4 Fig A).


610 P2 assistance increased parameters from about 2.593 to 2.653 million and computation


611 from 6.45 to 7.24 GFLOPs without improving accuracy, whereas P3 CBAM added negligible


612 computational cost but still did not surpass A0. All four checkpoints passed FP32 ONNX


613 export/runtime-parity checks, and the test partition remained unused during screening. Attention


614 and multi-scale modules can improve other agricultural tasks &#91;59,60&#93;, but their value is task- and


615 placement-dependent. Here, the negative ablation result indicates that the residual RRWIT


616 errors were not resolved by these three lightweight modifications and supports retaining the


617 simpler baseline for the formal benchmark (S4 Fig B).


618 The P2-assisted pathway produced the clearest negative result. Relative to A0, A1 reduced


619 validation mAP50-95 by 0.112 and mAP50 by 0.124 while increasing theoretical computation by


620 about 12.2%. A2 was much closer to baseline, reducing mAP50-95 by only 0.013 with negligible


621 computational overhead, showing that the result should not be generalized to attention


622 mechanisms as a class. A3 illustrates why a single operating-point statistic should not


623 determine model selection; its precision increased slightly to 0.814, but recall fell to 0.685 and


624 both mAP metrics declined. The prespecified mAP50-95 criterion therefore retained A0 (S4 Figs


625 A and B).


626
## **3.5 Cross-framework benchmarking confirms**


627
## **competitive RRWIT performance while newer YOLO**


628
## **architectures improve localization accuracy**


629 The controlled benchmark compared eight detector configurations under identical source

630 image partitions and frozen validation-to-test procedures. The architectures, implementations,


631 reference input sizes, and published model characteristics are summarized in S1 Table. This


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.17.767224; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.17.767224)

(which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC
105 and is also made available for use under a CC0 license.


632 breadth was intended to compare RRWIT not only with preceding and subsequent YOLO


633 generations but also with one-stage mobile, compound-scaled, and two-stage detector designs.


634 The validation-stage analysis established the framework ranking before locked-test evaluation


635 (S5 Fig). Across seeds 42, 123, and 2025, YOLO26n achieved the highest validation mAP50–


636 95 (0.834 ± 0.006), followed by YOLO12n (0.819 ± 0.006) and RRWIT (0.808 ± 0.009).


637 YOLOv8n, EfficientDet-D0, and YOLOv5nu followed at 0.790 ± 0.009, 0.776 ± 0.006, and 0.775


638 ± 0.007, respectively, whereas Faster R-CNN-320 and SSDLite320 produced substantially


639 lower values of 0.673 ± 0.007 and 0.648 ± 0.005. Validation mAP50 was high for the YOLO

640 family models, reaching 0.988 ± 0.003 for YOLO26n and 0.971 ± 0.007 for RRWIT. Macro and


641 support-weighted F1 scores (S5 Fig C) showed a similar high-performing group, although their


642 ordering was not identical to mAP50–95 because these statistics reflect a fixed confidence


643 operating point rather than integration over the precision-recall curve.


644 The validation analysis also revealed substantial dependence on object scale (S5 Fig D).


645 Using the standard COCO object-size categories employed in the benchmark evaluation, AP for


646 large objects was generally higher than for medium and especially small objects across


647 frameworks, although the magnitude of this difference varied among architectures. This pattern


648 is consistent with the well-established difficulty of detecting small objects, which contain fewer


649 pixels and provide less spatial and feature information for localization and classification &#91;15,17&#93;.


650 Importantly, this benchmark result is not directly comparable with the non-monotonic size


651 pattern observed in the seed-0 RRWIT audit in Section 3.3. That analysis used RRWIT-specific


652 relative-area categories, including a separate tiny-object class, and reported the proportion of


653 images that were both correctly classified and localized for a single historical checkpoint,


654 whereas S5 Fig D reports COCO-style size-stratified AP across independently trained


655 benchmark models. YOLO26n showed particularly strong performance across the three COCO


656 scale strata, supporting its validation-stage performance. The per-species comparison between


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.17.767224; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.17.767224)

(which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC
105 and is also made available for use under a CC0 license.


657 RRWIT and YOLO26n further indicated that the aggregate advantage of YOLO26n was


658 distributed across multiple taxa rather than being attributable to only one species (S5 Fig E, F).


659 Held-out test partitions subset results broadly preserved validation stage performance (Fig


660 9). YOLO26n had the highest mAP50-95, 0.810 ± 0.006, followed by YOLO12n at 0.804 ± 0.012


661 and RRWIT at 0.768 ± 0.004. RRWIT therefore ranked third of eight on the primary metric and


662 remained close to EfficientDet-D0 (0.758 ± 0.007) and YOLOv8n (0.757 ± 0.010), while


663 exceeding YOLOv5nu, Faster R-CNN-320, and SSDLite320 (Fig 9A Table 1).


664 **Fig 9. Held-out test partitions subset performance of eight object-detection**


665 **frameworks under the frozen three-seed benchmark.** **(A)** Test mAP50–95 and **(B)** test


666 mAP50 across seeds 42, 123, and 2025. Blue circles indicate individual seed runs; open


667 diamonds and horizontal error bars indicate arithmetic mean ± sample SD. **(C)** Macro and


668 support-weighted F1 calculated using confidence thresholds selected exclusively on the


669 validation partition and frozen before test evaluation. **(D)** Strict-localization performance


670 summarized by mAP75 and average recall at a maximum of 100 detections per image (AR100).


671 **(E)** Per-species test AP50–95 and **(F)** per-species test F1 comparing RRWIT (YOLO11n) with


672 the highest-ranked framework, YOLO26n. Horizontal error bars in panels E and F represent


673 sample SD across the three seeds; gray connecting lines aid within-species comparison, and _n_


674 denotes the number of test annotations for each species.


675 **Table 1. Predictive performance and computational footprint of the eight trained object-detection frameworks under the frozen**


676 **RRWIT benchmark.**

|Framework/&lt;br&gt;model|Input(px)|Params(M)|GFLOPs|FP32param&lt;br&gt;.&lt;br&gt;memory(MiB)|Val mAP50&lt;br&gt;.|Val mAP50 -95&lt;br&gt;.|TestmAP50|Test mAP50 -95|
|---|---|---|---|---|---|---|---|---|
|YOLOv5nu|640 x 640|2.511|7.078|9.58|0.971 ± 0.005|0.775 ± 0.007|0.947 ± 0.013|0.743 ± 0.011|
|YOLOv8n|640 x 640|3.014|8.097|11.50|0.968 ± 0.004|0.790 ± 0.009|0.944 ± 0.006|0.757 ± 0.010|
|RRWIT&lt;br&gt;(YOLO11n)|640 x 640|2.593|6.386|9.89|0.971 ± 0.007|0.808 ± 0.009|0.938 ± 0.005|0.768 ± 0.004|
|YOLO12n|640 x 640|2.571|7.314|9.81|0.968 ± 0.006|0.819 ± 0.006|0.962 ± 0.009|0.804 ± 0.012|
|YOLO26n|640 x 640|2.509|5.768|9.57|0.988 ± 0.003|0.834 ± 0.006|0.976 ± 0.010|0.810 ± 0.006|
|SSDLite320&lt;br&gt;MobileNetV3-&lt;br&gt;Large|320 x 320|2.387|0.863|9.10|0.885 ± 0.011|0.648 ± 0.005|0.879 ± 0.012|0.658 ± 0.008|
|EfficientDet-D0|512 x 512|3.836|4.507|14.63|0.962 ± 0.008|0.776 ± 0.006|0.939 ± 0.010|0.758 ± 0.007|
|Faster R-CNN&lt;br&gt;MobileNetV3-&lt;br&gt;Large-320 FPN|320 x 320|19.021|1.438|72.56|0.924 ± 0.012|0.673 ± 0.007|0.907 ± 0.009|0.659 ± 0.004|



677 Validation and locked-test metrics are arithmetic mean ± sample SD across seeds 42, 123, and 2025. Checkpoint and confidence-threshold


678 selection used validation data only; locked-test results were not used for model selection. GFLOPs are theoretical batch-1 operation counts at


679 the frozen input size and are not measured latency. FP32 parameter memory represents registered 32-bit parameter tensors and is not


680 checkpoint or exported-model file size. mAP50 is mean average precision at IoU 0.50; mAP50-95 is COCO-style mean average precision


681 averaged over IoU thresholds 0.50:0.05:0.95.


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.17.767224; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.17.767224)

(which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC
105 and is also made available for use under a CC0 license.


682 Metric dependence refined the comparison between models (Fig 9). RRWIT test mAP50


683 was 0.938 ± 0.005, compared with 0.976 ± 0.010 for YOLO26n and 0.962 ± 0.009 for


684 YOLO12n; YOLOv5nu, YOLOv8n, and EfficientDet-D0 were numerically similar to RRWIT at


685 IoU 0.50 despite lower mAP50-95 (Fig 9B). At IoU 0.75, RRWIT reached 0.844 ± 0.002 versus


686 0.896 ± 0.010 for YOLO26n and 0.894 ± 0.016 for YOLO12n (Fig 9D). This pattern indicates


687 that the newer YOLO architectures gained primarily under stricter localization requirements.


688 RRWIT nevertheless retained high support-weighted F1 (0.977 ± 0.001) and low between-seed


689 mAP50-95 SD (0.004) (Fig 9C).


690 Average recall at a maximum of 100 detections per image (AR100) showed a related


691 pattern: YOLO26n reached 0.854 ± 0.006, YOLO12n and EfficientDet-D0 approximately 0.830,


692 and RRWIT 0.796 ± 0.004 (Fig 9D). At validation-selected confidence thresholds, macro F1 was


693 0.956 ± 0.014 for YOLO12n, 0.954 ± 0.010 for YOLO26n, and 0.939 ± 0.005 for RRWIT (Fig


694 9C). These operating-point F1 values and COCO-style AP are complementary rather than


695 interchangeable: F1 summarizes one frozen threshold, whereas AP integrates ranked


696 predictions and, for mAP50-95, multiple localization thresholds.


697 Validation-to-test changes also varied across frameworks. RRWIT changed from 0.808 ±


698 0.009 validation mAP50-95 to 0.768 ± 0.004 on locked test (approximately -0.040), compared


699 with changes of approximately -0.024 for YOLO26n and -0.015 for YOLO12n (Table 1). The


700 broad ranking nevertheless remained stable, providing a useful internal generalization check.


701 The benchmark also places the RRWIT model choice in a temporal context. YOLO11n was


702 a compact current-generation detector when RRWIT development was designed, whereas


703 YOLO12 (attention-centric real-time detector design) and YOLO26 (introduced a newer


704 architecture incorporating a lighter regression head and native end-to-end inference capabilities)


705 became available later &#91;22,41,42&#93;. Similar within-dataset agricultural comparisons increasingly


706 showed that the model with the highest accuracy is not necessarily the model with the preferred


707 deployment trade-off &#91;61,62&#93;. In the present dataset, YOLO26n improved strict-IoU accuracy


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.17.767224; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.17.767224)

(which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC
105 and is also made available for use under a CC0 license.


708 while using slightly fewer parameters and lower theoretical computation than RRWIT, making it


709 the strongest evaluated candidate for future RRWIT development without implying that its


710 ranking will remain unchanged as our biological dataset and computer vision architectures


711 evolves. The present benchmark therefore should not be interpreted as evidence that the initial


712 YOLO11n choice was inappropriate; rather, it demonstrates how the performance frontier


713 changed as newer architectures became available.


714


715
## **3.6 Species-level performance identifies persistent**


716
## **challenges among visually similar and**


717
## **underrepresented taxa**


718 Species-level evaluation revealed substantially greater heterogeneity than was apparent


719 from framework-level summary metrics. Across the three prespecified seeds, RRWIT showed


720 strong locked-test performance for several taxa, including _Conotrachelus naso_ and _Anthonomus_


721 _grandis grandis_, whereas lower AP50–95 occurred for a subset of taxa including _Curculio fulvus_,


722 _Anthonomus hunteri_, _Anthonomus eugenii_, and _Curculio pardalis_ . Importantly, test


723 representation varied markedly among species, from 595 annotated _A. grandis grandis_ and 223


724 _A. fulvus_ objects to only 22 _C. caryae_, 13 _A. hunteri_, 10 _C. pardalis_, and four _C. fulvus_ .


725 Consequently, the larger between-seed uncertainty observed for some sparsely represented


726 taxa, particularly _C. fulvus_ and _C. pardalis_, should be interpreted in the context of their limited


727 test support rather than as precise estimates of population-level performance (Fig 10A).


728 **Fig 10. Species-level locked-test performance of RRWIT across three independent**


729 **training seeds.** (A) Species-specific AP50-95 and (B) F1 at validation-selected confidence


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.17.767224; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.17.767224)

(which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC
105 and is also made available for use under a CC0 license.


730 thresholds. Points are arithmetic means and horizontal error bars are sample SD; numbers in


731 parentheses on y-axis show locked-test annotation support.


732 AP50–95 (Fig 10A) and F1 (Fig 10B) also provided complementary information about


733 the origin of species-level errors. Some classes with comparatively low AP50–95 nevertheless


734 maintained high F1 scores at the validation-selected operating threshold (Fig 10). This was


735 particularly evident for _A. hunteri_, indicating that successful detection and class assignment at


736 IoU 0.50 can coexist with reduced localization consistency when performance is averaged over


737 progressively stricter IoU thresholds from 0.50 to 0.95. Conversely, _C. fulvus_ and _C. pardalis_


738 showed reductions in both AP50–95 and F1, although interpretation is necessarily constrained


739 by their very small test supports. Thus, low species-level AP should not automatically be


740 interpreted as a pure classification failure; it can reflect some combination of localization


741 precision, missed detections, species substitution, and threshold-dependent classification


742 performance.


743 The three-seed confusion matrices reinforced this pattern (Figs 11, 12 and S6) with most


744 classes remaining strongly diagonal. The mean on the held-out previously unseen test partitions


745 subset diagonal proportions were 1.00 for _A. grandis grandis_, 0.99 for _A. peninsularis_ and _A._


746 _texanus_, 0.98 for _A. fulvus_ and _C. naso_, and 0.95 for _A. eugenii_ . _Curculio_ taxa with lower


747 number of images were more heterogeneous, with corresponding means of approximately 0.91


748 for _C. caryae_, 0.75 for _C. fulvus_, 0.87 for _C. pardalis_, and 0.90 for _C. proboscideus_, together with


749 more visible substitutions and unmatched-ground-truth proportions (Fig 11). Fine-grained


750 _Anthonomus_ and _Curculio_ identification depends on combinations of subtle morphological


751 characters &#91;49,50&#93;, providing a biological reason why some distinctions are difficult in two

752 dimensional images. Class abundance was not a complete explanation; however, some taxa


753 with lower number of images performed well, and some more represented taxa remained harder


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.17.767224; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.17.767224)

(which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC
105 and is also made available for use under a CC0 license.


754 than the best classes. Long-tailed support and high inter-/intra-class visual variation are


755 recognized challenges in insect computer vision &#91;63-65&#93;.


756 **Fig 11. Validation and held-out test partitions subset confusion patterns for RRWIT**


757 **(YOLO11n) across three independent training seeds.** Mean row-normalized confusion


758 matrices are shown for the validation set (A) and locked test set (B). Each cell is the arithmetic


759 mean across seeds 42, 123, and 2025 after independent row normalization. Rows are true


760 categories and columns are predicted categories. Off-diagonal biological cells indicate species


761 substitutions, the Background column represents unmatched ground-truth objects (false


762 negatives), and the Background row represents unmatched predictions (false positives) by


763 predicted class. The Background-Background cell is undefined and left blank. Matching and


764 thresholding followed the frozen evaluation procedure.


765


766 The difficult classes were not unique to RRWIT. YOLO26n showed a broadly similar


767 taxonomic error structure (Fig 12), and YOLO12n showed the same persistent difficulty for


768 several low-support _Curculio_ taxa (S6 Fig). For YOLO26n, the mean held-out previously unseen


769 test partitions subset diagonal proportions for _C. fulvus_, _C. pardalis_, and _C. proboscideus_ were


770 approximately 0.83, 0.83, and 0.90, respectively. YOLO12n showed approximately 0.83, 0.90,


771 and 0.95 for the same taxa (Fig 12; S6 Fig). Cross-framework persistence suggests that a


772 substantial component of residual error reflects task/data properties, limited representation,


773 subtle morphology, orientation, and visibility of diagnostic structures rather than a single RRWIT


774 architecture. Related insect-monitoring studies similarly emphasize field-representative


775 retraining and the difficulty of species-level discrimination among morphologically uniform


776 groups &#91;20,66&#93;.


777 **Fig 12. Validation and held-out test partitions subset confusion patterns for**


778 **YOLO26n across three independent seeds.** Mean row-normalized confusion matrices are


779 shown for the validation set (A) and locked test set (B). Each cell is the arithmetic mean across


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.17.767224; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.17.767224)

(which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC
105 and is also made available for use under a CC0 license.


780 seeds 42, 123, and 2025 after independent row normalization. Rows are true categories and


781 columns are predicted categories. Off-diagonal biological cells indicate species substitutions,


782 the Background column represents unmatched ground-truth objects (false negatives), and the


783 Background row represents unmatched predictions (false positives) by predicted class. The


784 Background-Background cell is undefined and left blank. Matching and thresholding followed


785 the frozen evaluation procedure.


786


787 The combination of Fig 10 and the multi-framework confusion matrices indicates that future


788 improvements should prioritize targeted acquisition of diverse images for difficult taxa,


789 especially low-support _Curculio_ species, while deliberately varying specimen orientation, scale,


790 illumination, background, sex where relevant, and visibility of diagnostic morphological


791 structures. Such targeted enrichment would be more informative than indiscriminately


792 increasing already well-represented classes. From an applied surveillance perspective, these


793 species-level results also define where human verification remains most valuable. Automated


794 inference can substantially reduce the burden of screening large numbers of trap images, but


795 predictions involving low-support or repeatedly confused taxa should be treated with greater


796 caution than predictions for consistently high-performing classes. This is particularly important


797 when species identity would trigger a management or regulatory response. Accordingly, RRWIT


798 is better viewed as a decision-support system for rapid taxonomic screening than as a


799 replacement for expert confirmation in the most difficult cases. The three-seed analyses make


800 that limitation explicit while also showing that the majority of species are detected with highly


801 consistent class-level performance.


802


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.17.767224; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.17.767224)

(which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC
105 and is also made available for use under a CC0 license.


803
## **3.7 Accuracy-complexity trade-offs favor lightweight**


804
## **one-stage detectors**


805 The leading models occupied a narrow lightweight regime rather than obtaining higher


806 accuracy through substantially greater capacity (Table 1). YOLO26n combined test mAP50-95


807 of 0.810 ± 0.006 with 2.509 million parameters, 5.768 GFLOPs, and 9.57 MiB of FP32


808 parameter memory; YOLO12n used 2.571 million parameters and 7.314 GFLOPs; RRWIT used


809 2.593 million and 6.386 GFLOPs. By contrast, SSDLite320 had only 0.863 GFLOPs but lower


810 mAP50-95 (0.658 ± 0.008), while Faster R-CNN-320 contained 19.021 million parameters and


811 72.56 MiB of FP32 parameter memory yet reached 0.659 ± 0.004.


812 These observations illustrate why parameter count and theoretical operation count should


813 not be interpreted as runtime. The models also used different frozen input sizes: the YOLO

814 family detectors used 640 × 640 pixels, EfficientDet-D0 512 × 512, and the SSDLite/Faster R

815 CNN configurations 320 × 320; therefore, actual speed depends on operator structure, memory


816 access, inference-engine implementation, preprocessing, post-processing, and hardware


817 utilization &#91;67&#93;. Recent agricultural work similarly treats accuracy, model size, and computation


818 as joint design objectives rather than optimizing a single scalar &#91;68-70&#93;. For this weevil task,


819 compact YOLO-family detectors occupied the most favorable accuracy-complexity region


820 among the evaluated configurations; actual deployment performance required direct runtime


821 measurement.


822 Within these constraints, the strongest result is not that all one-stage detectors are


823 universally superior, but that the compact YOLO-family models occupied the most favorable


824 accuracy-complexity region for this particular fine-grained task. Similar parameter budgets did


825 not yield equivalent localization performance, and YOLO26n improved the RRWIT accuracy

826 complexity frontier without moving to a substantially larger network (Table 1). This makes


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.17.767224; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.17.767224)

(which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC
105 and is also made available for use under a CC0 license.


827 lightweight one-stage detection a defensible design space for future RRWIT development while


828 leaving actual device performance to direct deployment measurements.


829
## **3.8 INT8 quantization substantially reduces model**


830
## **size but does not guarantee faster CPU inference**


831 Common static INT8 quantization substantially compressed all four deployment candidates,


832 but accuracy retention was architecture-dependent (Table 2). YOLO26n, YOLO12n, and RRWIT


833 retained 98.6%, 98.8%, and 98.1% of FP32 validation mAP50-95, respectively, while their


834 serialized ONNX artifacts shrank by approximately 69-71%. RRWIT decreased from 10.09 to


835 2.91 MiB (71.2%). Under the prespecified aggregate gate, only YOLO12n met all three accuracy


836 tolerances; YOLO26n and RRWIT exceeded the mAP50-95 loss threshold slightly while


837 retaining mAP50 and weighted F1 within tolerance. This is consistent with the storage benefit


838 expected from lower-precision integer representations &#91;71&#93;.


839 Table 2. Validation accuracy, serialized ONNX model size, and CPU inference characteristics of four validation-selected deployment candidates


840 after common INT8 post-training quantization.


































|Model|Seed|mAP50&lt;br&gt;FP32→INT8|mAP50 -95&lt;br&gt;FP32→INT8|AP5095&lt;br&gt;-&lt;br&gt;retained(%)|WeightedF1&lt;br&gt;FP32→INT8|ONNXsize&lt;br&gt;FP32→INT8&lt;br&gt;(MiB)|Size↓(%)|CPUE2E&lt;br&gt;FP32→INT8(ms)|INT8&lt;br&gt;p95(ms)|INT8img/s|Outcome|
|---|---|---|---|---|---|---|---|---|---|---|---|
|YOLO26n|123|0.987→0.988|0.839→0.828|98.6|0.983→0.981|9.32→2.79|70.1|45.16→82.35|88.38|12.01|Minor&lt;br&gt;mAP50-95&lt;br&gt;loss|
|YOLO12n|42|0.974→0.976|0.825→0.816|98.8|0.986→0.979|10.04→3.07|69.4|64.06→107.67|112.00|9.25|Qualified|
|RRWIT&lt;br&gt;(YOLO11n)|2025|0.978→0.971|0.818→0.802|98.1|0.985→0.983|10.09→2.91|71.2|60.81→82.63|88.23|11.99|Minor&lt;br&gt;mAP50-95&lt;br&gt;loss|
|EfficientDet-&lt;br&gt;D0|42|0.971→0.283|0.783→0.165|21.1|0.959→0.379|15.50→5.04|67.5|135.58→227.13|260.97|4.34|Further&lt;br&gt;optimization&lt;br&gt;needed|



841 Each row is one validation-selected checkpoint, not a three-seed mean. Deployment qualification used validation data only. The common INT8


842 protocol used ONNX Runtime static QDQ quantization with MinMax calibration, signed 8-bit activations and weights, per-tensor weight


843 quantization, and 512 frozen training images for calibration. CPU end-to-end (E2E) latency is the batch-1 median including image decoding,


844 family-specific preprocessing, ONNX Runtime inference, detector-specific decoding/non-maximum suppression, and canonicalization, measured


845 with four intra-operation threads on 128 frozen training images after 16 warm-up images and over three timed rounds. The prespecified accuracy


846 gate required absolute declines ≤0.010 in mAP50-95, mAP50, and weighted F1.


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.17.767224; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.17.767224)

(which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC
105 and is also made available for use under a CC0 license.


847 Compression did not accelerate the tested CPU execution path. Median end-to-end latency


848 increased from 45.16 to 82.35 ms/image for YOLO26n, 64.06 to 107.67 for YOLO12n, 60.81 to


849 82.63 for RRWIT, and 135.58 to 227.13 for EfficientDet-D0. Integer inference can be faster on


850 software/hardware stacks optimized for low-precision kernels, but quantization format alone


851 does not guarantee speed. Hardware-targeted deployment studies illustrate that device-specific


852 runtimes can realize acceleration not observed under a generic CPU path &#91;72,73&#93;.


853 This result is compatible with hardware-targeted quantization studies rather than


854 contradictory to them. Wang et al. converted YOLOv11 to an INT8 RKNN representation


855 optimized for an RK3568 edge device and reported a size reduction from 10,156 to 4,194 kB


856 together with a single-image detection-time reduction from 52.1 to 33.6 ms &#91;72&#93;. Chen et al.


857 likewise evaluated a YOLO11-derived model on a Jetson TX2 under INT8 CPU-only


858 deployment and reported strong accuracy retention &#91;73&#93;. Their results illustrate the benefit when


859 graph conversion, integer kernels, and hardware are co-optimized; our experiment instead


860 tested a common ONNX Runtime CPU route across architectures.


861 EfficientDet-D0 was particularly sensitive to the common quantization protocol. For


862 example, mAP50-95 fell from 0.783 to 0.165 and weighted F1 from 0.959 to 0.379 despite a


863 67.5% size reduction (Tables 2 and 3). Per-channel quantization substantially recovered


864 accuracy, with U8U8/per-channel reaching mAP50-95 of 0.556 and weighted F1 of 0.800, but


865 no diagnostic configuration restored FP32-equivalent accuracy or latency. A bridge-control


866 analysis reproduced the calibration bytes, preprocessing stream, quantization parameters, and


867 common artifacts, excluding artifact drift as an explanation for the collapse. Thus, quantization


868 should be treated as an architecture- and hardware-specific optimization problem rather than a


869 universal conversion step.


870 Table 3. Architecture-specific EfficientDet-D0 INT8 quantization diagnostics following substantial accuracy degradation under the common


871 quantization protocol used in Table 2.













|Variant|Role|mAP50|mAP50 -95|WeightedF1|ΔmAP50 -95 vs&lt;br&gt;FP32|ONNXsize(MiB)|CPUE2Emedian&lt;br&gt;(ms)|Throughput&lt;br&gt;(img/s)|
|---|---|---|---|---|---|---|---|---|
|FP32|Reference|0.971|0.783|0.959|0.000|15.50|135.58|7.36|
|S8S8/PT|Common protocol|0.283|0.165|0.379|−0.618|5.04|227.13|4.34|
|S8S8/PC|Diagnostic only|0.768|0.544|0.789|−0.238|5.47|205.78|4.79|
|U8U8/PT|Diagnostic only|0.415|0.266|0.526|−0.517|5.04|182.32|5.38|
|U8U8/PC|Diagnostic only|0.785|0.556|0.800|−0.227|5.47|166.48|5.94|


872 FP32 is the reference and S8S8/PT is the authoritative common cross-framework INT8 protocol. S8S8/PC, U8U8/PT, and U8U8/PC were


873 architecture-specific diagnostics only. S8S8 and U8U8 denote signed and unsigned 8-bit activation/weight formats; PT and PC denote per-tensor


874 and per-channel weight quantization. Validation metrics used the same 1,690-image validation set; the locked test set remained prohibited. CPU


875 E2E timing followed the controlled batch-1 ONNX Runtime protocol used for Table 2.


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.17.767224; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.17.767224)

(which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC
105 and is also made available for use under a CC0 license.


876 Deployment experiments therefore separate four properties (theoretical computation,


877 serialized storage, predictive retention after quantization, and measured runtime) that are often


878 conflated. RRWIT can be compressed to a roughly 3-MiB INT8 artifact while retaining high


879 validation accuracy, but speed claims require backend- and device-specific testing. For RRWIT


880 specifically, INT8 therefore provides a viable route to substantially lower serialized storage, but


881 additional backend- and hardware-specific optimization is required before claiming a speed


882 advantage. The current cloud-hosted prototype supports both image upload and live capture


883 and returns bounding boxes, species labels, and confidence scores through an interactive


884 interface (S1 Video). It currently operates in a CPU-based cloud environment (2 vCPUs and 16


885 GB RAM). Application-level latency and scalability were not systematically benchmarked, so


886 apparent web-response time is not reported as a deployment metric. Other metrics such as


887 energy use, and representative mobile/edge performance also remain to be evaluated.


888
## **3.9 Implications, limitations, and future development** 889 of RRWIT


890 Taken together, the experiments support the original design premise without implying that


891 YOLO11n is universally optimal. Pretrained initialization improved mean performance and


892 reduced between-seed variability. Quantitative EigenGradCAM showed that attribution


893 localization is strongly scale-dependent and that object-focused attribution does not ensure


894 correct species classification. Furthermore, the targeted A1-A3 modifications failed to improve


895 the baseline and the frozen benchmark showed that newer YOLO12n and especially YOLO26n


896 can improve strict-IoU localization at a comparable lightweight footprint. RRWIT is therefore


897 best understood as an evolving identification framework, with YOLO26n representing a strong


898 candidate backbone for future versions rather than a guaranteed permanent successor.


899 Confusion patterns further showed that the remaining classification errors were not


900 distributed uniformly across taxa (Fig 11). RRWIT retained particularly strong discrimination of


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.17.767224; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.17.767224)

(which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC
105 and is also made available for use under a CC0 license.


901 the boll weevil, _A. grandis grandis_, for which the mean diagonal proportion across the three


902 independently trained models was 99.66% on the locked test set. Residual substitutions from


903 this pest species were rare and individually below the 2% annotation threshold used in Fig 11.


904 In contrast, the pecan weevil, _C. caryae_, remained more challenging: its mean test diagonal was


905 90.91%, with 4.55% assigned to _C. proboscideus_, 1.52% to _A. fulvus_, and 3.03% unmatched as


906 background. Importantly, most matched wrong-species assignments across the test evaluations


907 occurred between non-pest taxa rather than across the pest/non-pest boundary, indicating that


908 a substantial component of the remaining species-level error involved discrimination among


909 morphologically similar non-target weevils rather than systematic confusion of the two target


910 pest species (Fig 11).


911 The newer YOLO26n architecture reduced several of these residual errors while preserving


912 the same general pattern (Fig 12). Its mean locked-test diagonal proportions reached 99.72%


913 for _A. grandis grandis_ and 95.45% for _C. caryae_ . For pecan weevil, the principal residual


914 substitution was again _C. proboscideus_ (3.03%), with 1.52% assigned to background. Across


915 the three test evaluations, the great majority of YOLO26n matched wrong-species assignments


916 likewise occurred among non-pest taxa. Thus, the confusion analysis identifies two


917 complementary features of the benchmark: boll-weevil discrimination was highly stable across


918 both detectors, whereas the more persistent errors were concentrated among particular


919 morphologically similar taxa, especially within the more difficult _Curculio_ group (Figs 11 and 12).


920 These results support targeted expansion of underrepresented and difficult species in future


921 RRWIT datasets rather than relying solely on architectural improvements to resolve the


922 remaining taxonomic ambiguity.


923


924 The current internal test partition preserves image sources and historical acquisition


925 contexts represented in the overall dataset. It is not intended to serve as external validation from


926 new locations, institutions, trapping programs, devices, dates, or end users. Insect-monitoring


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.17.767224; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.17.767224)

(which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC
105 and is also made available for use under a CC0 license.


927 research emphasizes that incomplete, imbalanced reference libraries and changing


928 environmental context can drive errors, including assignment of unfamiliar taxa to visually


929 similar known classes &#91;74&#93;. Future RRWIT upgrades will therefore prioritize improving


930 annotation from bounding boxes to instance segmentation, keypoints detections, and Panoptic


931 segmentation. It will also include diverse new images of low-support and repeatedly confused


932 taxa, independent external testing, and mechanisms for flagging unfamiliar or uncertain


933 specimens rather than assuming every insect belongs to one of the 14 trained classes. An


934 expert-in-the-loop cycle in which such specimens are reviewed, annotated, and incorporated


935 into later training releases would allow the biological reference library to expand while retaining


936 taxonomic oversight &#91;74&#93;.


937 As our image collection expands, the scientific goals of RRWIT can move beyond simply


938 localizing a weevil and identifying its taxon toward investigating the specific morphological


939 structures that inform those predictions. Future enhancements could incorporate instance


940 segmentation to isolate complete insect silhouettes and anatomical regions, and utilize


941 taxonomist-defined keypoints to encode homologous landmarks relevant to diagnostic features.


942 Mask R-CNN introduced a unified framework for object detection and instance segmentation


943 and can be extended to estimate keypoints &#91;75&#93;, while DeepPoseKit demonstrated fast and


944 robust dense keypoint estimation for insects and other animals &#91;76&#93;. Integrating these richer


945 annotation types would enable direct, biologically meaningful comparisons between expert

946 defined taxonomic characters and attribution maps derived using techniques such as


947 EigenGradCAM particularly in complex images containing multiple insects, debris, vegetation,


948 and trap components.


949


950 Whole-trap or field imagery may eventually justify scene-level methods. Panoptic


951 segmentation unifies individual-object and semantic scene interpretation &#91;77&#93;, and PhenoBench


952 demonstrates coordinated semantic, instance, and hierarchical agricultural image tasks,


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.17.767224; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.17.767224)

(which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC
105 and is also made available for use under a CC0 license.


953 including evaluation on an unseen field &#91;78&#93;. In parallel, direct benchmarking on intended ARM


954 CPUs, neural-processing units, or other accelerators is required before on-device performance


955 can be claimed.


956 The prototype also contains an exploratory conversational artificial intelligence (AI)


957 assistant (S1 Video), illustrating a longer-term path toward combining visual identification with


958 curated taxonomic, biological, surveillance, and management information. That component was


959 not systematically evaluated and is not part of the detection benchmark. Any future decision

960 support implementation would require separate assessment of retrieval accuracy, factual


961 reliability, uncertainty communication, provenance, usability, and safeguards against


962 unsupported responses. The current study therefore establishes a technically evaluated visual

963 identification foundation while keeping these broader capabilities explicitly prospective.

# 964 4 Conclusion


965 RRWIT demonstrates that transfer learning with a lightweight detector can support accurate


966 identification of morphologically similar pests and non-pest weevils. Replicated initialization


967 controls, quantitative EigenGradCAM, targeted ablation, and the frozen eight-framework


968 comparison provide a substantially stronger basis than a single-model evaluation. RRWIT


969 remained competitive, but YOLO26n achieved higher strict-IoU localization accuracy with a


970 comparable or smaller computational footprint (Table 1), showing that RRWIT should be treated


971 as an extensible framework rather than a system permanently tied to YOLO11n.


972 The results also identify two priorities for future development. First, species-level errors and


973 low support for several taxa make representative biological data and independent external


974 validation central to further gains. Second, model compactness is not equivalent to execution


975 speed. INT8 quantization reduced RRWIT from 10.09 to 2.91 MiB but did not improve end-to

976 end computational latency on the tested CPU/ONNX Runtime path (Table 2). Deployment


977 claims should therefore be based on measurements from the intended hardware and runtime.


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.17.767224; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.17.767224)

(which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC
105 and is also made available for use under a CC0 license.


978 With continued expert-verified image acquisition, external validation, and hardware-specific


979 optimization, RRWIT provides a practical foundation for progressively richer automated weevil


980 identification. Future work can evaluate newer lightweight backbones and, as data permit,


981 instance segmentation, anatomically defined keypoints, and complex-scene interpretation while


982 retaining expert oversight for uncertain or management-critical cases. The existing web


983 prototype demonstrates an accessible interaction pathway, but operational field deployment and


984 the exploratory conversational component require separate validation.


985


986
# **Acknowledgements**


987 We thank the Texas Boll Weevil Eradication Foundation, Texas Department of Agriculture,


988 and the Texas Pecan Growers Association for field support and access to traps. We also thank


989 the Texas A&amp;M University Insect Collection for providing reference specimens. Special thanks


990 to Jayda Arriaga and Kiley Stout for help with imaging and insect pinning. We would like to


991 thank the ARS SCINet Office in general, and specifically Drs. Brian Stucky and Heather Savoy,


992 for providing general computer vision training for our team as well as Joey Jones for his efforts


993 in allocating GPU resources when they were most needed. Mention of trade names,


994 companies, or commercial products in this publication is solely for the purpose of


995 providing specific information and does not imply recommendation or endorsement of


996 the products by the U.S. Department of Agriculture. The USDA is an equal opportunity


997 provider and employer.


998


999


1000


1001


1002


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.17.767224; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.17.767224)

(which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC
105 and is also made available for use under a CC0 license.


1003 **References**


1004 &#91;1&#93; Oberprieler RG, Marvaldi AE, Anderson RS. Weevils, weevils, weevils everywhere. Zootaxa.


1005 2007;1668(1):491-520. doi:10.11646/zootaxa.1668.1.24.


1006 &#91;2&#93; McKenna DD, Sequeira AS, Marvaldi AE, Farrell BD. Temporal lags and overlap in the


1007 diversification of weevils and flowering plants. Proc Natl Acad Sci U S A.


1008 2009;106(17):7083-7088. doi:10.1073/pnas.0810618106.


1009 &#91;3&#93; Perkin LC, Cohen ZP, Sim SB, Geib SM, Childers AK, Smith TPL, et al. A chromosome level


1010 reference genome for the pecan weevil, Curculio caryae. Sci Data. 2026;13:706.


1011 doi:10.1038/s41597-026-07030-8.


1012 &#91;4&#93; Harris MK, Jackman JA. Pecan arthropod management. In: Wood BW, Payne JA, editors.


1013 Pecan husbandry: challenges and opportunities. First National Pecan Workshop


1014 Proceedings. Washington, DC: U.S. Department of Agriculture, Agricultural Research


1015 Service; 1991. ARS-96. pp. 6-15.


1016 &#91;5&#93; Harris M, Ree B, Cooper J, Jackman J, Young J, Lacewell R, et al. Economic impact of


1017 pecan integrated pest management implementation in Texas. J Econ Entomol.


1018 1998;91(5):1011-1020. doi:10.1093/jee/91.5.1011.


1019 &#91;6&#93; Mulder PG Jr, Harris MK, Grantham RA. Biology and management of the pecan weevil


1020 (Coleoptera: Curculionidae). J Integr Pest Manag. 2012;3(1):A1-A9. doi:10.1603/IPM10027.


1021 &#91;7&#93; Tedders WL, Wood BW. A new technique for monitoring pecan weevil emergence


1022 (Coleoptera: Curculionidae). J Entomol Sci. 1994;29(1):18-30.


1023 &#91;8&#93; Greenberg SM, Adamczyk JJ Jr, Armstrong JS. Principles and practices of integrated pest


1024 management on cotton in the Lower Rio Grande Valley of Texas. In: Larramendy ML,


1025 Soloneski S, editors. Integrated pest management and pest control-current and future


1026 tactics. Rijeka: InTech; 2012. pp. 3-34. doi:10.5772/31467.


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.17.767224; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.17.767224)

(which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC
105 and is also made available for use under a CC0 license.


1027 &#91;9&#93; U.S. Department of Agriculture, Animal and Plant Health Inspection Service. Cotton pests


1028 &#91;Internet&#93;. Riverdale (MD): USDA APHIS; &#91;cited 2026 Sep 5&#93;. Available from:


1029 https://www.aphis.usda.gov/plant-pests-diseases/cotton-pests


1030 &#91;10&#93; Alves AN, Souza WSR, Borges DL. Cotton pests classification in field-based images using


1031 deep residual networks. Comput Electron Agric. 2020;174:105488.


1032 doi:10.1016/j.compag.2020.105488.


1033 &#91;11&#93; Toscano-Miranda R, Aguilar J, Hoyos W, Caro M, Trebilcok A, Toro M. Different transfer


1034 learning approaches for insect pest classification in cotton. Appl Soft Comput.


1035 2024;153:111283. doi:10.1016/j.asoc.2024.111283.


1036 &#91;12&#93; Spiesman BJ, Gratton C, Gratton E, Hines H. Deep learning for identifying bee species


1037 from images of wings and pinned specimens. PLoS ONE. 2024;19(5):e0303383.


1038 doi:10.1371/journal.pone.0303383.


1039 &#91;13&#93; Liu L, Ouyang W, Wang X, Fieguth P, Chen J, Liu X, et al. Deep learning for generic object


1040 detection: a survey. Int J Comput Vis. 2020;128:261-318. doi:10.1007/s11263-019-01247-4.


1041 &#91;14&#93; De Cesaro Junior T, Rieder R. Automatic identification of insects from digital images: a


1042 survey. Comput Electron Agric. 2020;178:105784. doi:10.1016/j.compag.2020.105784.


1043 &#91;15&#93; Wang QJ, Zhang SY, Dong SF, Zhang GC, Yang J, Li R, et al. Pest24: a large-scale very


1044 small object data set of agricultural pests for multi-target detection. Comput Electron Agric.


1045 2020;175:105585. doi:10.1016/j.compag.2020.105585.


1046 &#91;16&#93; Jiao L, Dong S, Zhang S, Xie C, Wang H. AF-RCNN: an anchor-free convolutional neural


1047 network for multi-categories agricultural pest detection. Comput Electron Agric.


1048 2020;174:105522. doi:10.1016/j.compag.2020.105522.


1049 &#91;17&#93; Zhang W, Huang H, Sun Y, Wu X. AgriPest-YOLO: a rapid light-trap agricultural pest


1050 detection method based on deep learning. Front Plant Sci. 2022;13:1079384.


1051 doi:10.3389/fpls.2022.1079384.


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.17.767224; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.17.767224)

(which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC
105 and is also made available for use under a CC0 license.


1052 &#91;18&#93; Tang Z, Lu J, Chen Z, Qi F, Zhang L. Improved Pest-YOLO: real-time pest detection based


1053 on efficient channel attention mechanism and transformer encoder. Ecol Inform.


1054 2023;78:102340. doi:10.1016/j.ecoinf.2023.102340.


1055 &#91;19&#93; Chen JW, Lin WJ, Cheng HJ, Hung CL, Lin CY, Chen SP. A smartphone-based application


1056 for scale pest detection using multiple-object detection methods. Electronics.


1057 2021;10(4):372. doi:10.3390/electronics10040372.


1058 &#91;20&#93; Sittinger M, Uhler J, Pink M, Herz A. Insect detect: an open-source DIY camera trap for


1059 automated insect monitoring. PLoS ONE. 2024;19(4):e0295474.


1060 doi:10.1371/journal.pone.0295474.


1061 &#91;21&#93; Mostafa V, Sozol MSS, Mondal MRH. LeafDet: a lightweight and interpretable deep


1062 learning framework for tomato leaf disease detection. PLoS ONE. 2026;21(5):e0349501.


1063 doi:10.1371/journal.pone.0349501.


1064 &#91;22&#93; Jocher G, Qiu J. Ultralytics YOLO11. Version 11.0.0 &#91;software&#93;. 2024. Available from:


1065 https://github.com/ultralytics/ultralytics


1066 &#91;23&#93; Quinn M. Weevils of Texas &#91;Internet&#93;. Texas Entomology; &#91;cited 2026 Sep 5&#93;. Available


1067 from: https://www.texasento.net/TXWeevils.html


1068 &#91;24&#93; Franz NM, Chamorro ML, Fleming D, Setliff GP, O’Brien CW. Weevils of North America


1069 (WoNA) &#91;dataset&#93;. Ag Data Commons; 2017 &#91;cited 2026 Sep 5&#93;. Available from:


1070 https://agdatacommons.nal.usda.gov/articles/dataset/Weevils_of_North_America_WoNA_/2


1071 4852279


1072 &#91;25&#93; iNaturalist. iNaturalist &#91;Internet&#93;. &#91;cited 2026 Sep 5&#93;. Available from:


1073 https://www.inaturalist.org/


1074 &#91;26&#93; BugGuide. Coleoptera: Curculionidae &#91;Internet&#93;. Iowa State University; &#91;cited 2026 Sep 5&#93;.


1075 Available from: https://bugguide.net/node/view/15740


1076 &#91;27&#93; Rife TW, Poland JA. Field Book: an open-source application for field data collection on


1077 Android. Crop Sci. 2014;54(4):1624-1627. doi:10.2135/cropsci2013.08.0579.


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.17.767224; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.17.767224)

(which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC
105 and is also made available for use under a CC0 license.


1078 &#91;28&#93; Clark JA, contributors. Pillow: Python Imaging Library (Fork) &#91;software&#93;. &#91;cited 2026 Sep 5&#93;.


1079 Available from: https://pillow.readthedocs.io/


1080 &#91;29&#93; Liu S, Zeng Z, Ren T, Li F, Zhang H, Yang J, et al. Grounding DINO: marrying DINO with


1081 grounded pre-training for open-set object detection. In: Computer Vision-ECCV 2024.


1082 Cham: Springer; 2025. pp. 38-55. doi:10.1007/978-3-031-72970-6_3.


1083 &#91;30&#93; Dwyer B, Nelson J, Hansen T, et al. Roboflow. Version 1.0 &#91;software&#93;. 2026. Available from:


1084 https://roboflow.com/


1085 &#91;31&#93; Lin TY, Maire M, Belongie S, Hays J, Perona P, Ramanan D, et al. Microsoft COCO:


1086 common objects in context. In: Fleet D, Pajdla T, Schiele B, Tuytelaars T, editors. Computer


1087 Vision-ECCV 2014. Cham: Springer; 2014. pp. 740-755. doi:10.1007/978-3-319-10602

1088 1_48.


1089 &#91;32&#93; Abid A, Abdalla A, Abid A, Khan D, Alfozan A, Zou J. Gradio: hassle-free sharing and


1090 testing of ML models in the wild. arXiv &#91;Preprint&#93;. 2019. arXiv:1906.02569.


1091 &#91;33&#93; Hugging Face. Hugging Face Spaces &#91;Internet&#93;. &#91;cited 2026 Sep 5&#93;. Available from:


1092 https://huggingface.co/spaces


1093 &#91;34&#93; Gildenblat J, contributors. PyTorch library for CAM methods &#91;software&#93;. GitHub; 2021.


1094 Available from: https://github.com/jacobgil/pytorch-grad-cam


1095 &#91;35&#93; Selvaraju RR, Cogswell M, Das A, Vedantam R, Parikh D, Batra D. Grad-CAM: visual


1096 explanations from deep networks via gradient-based localization. In: 2017 IEEE


1097 International Conference on Computer Vision. 2017. pp. 618-626.


1098 doi:10.1109/ICCV.2017.74.


1099 &#91;36&#93; Muhammad MB, Yeasin M. Eigen-CAM: Class activation map using principal components.


1100 In: 2020 International Joint Conference on Neural Networks. 2020. pp. 1-7.


1101 doi:10.1109/IJCNN48605.2020.9206626.


1102 &#91;37&#93; Adebayo J, Gilmer J, Muelly M, Goodfellow I, Hardt M, Kim B. Sanity checks for saliency


1103 maps. Adv Neural Inf Process Syst. 2018;31.


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.17.767224; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.17.767224)

(which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC
105 and is also made available for use under a CC0 license.


1104 &#91;38&#93; Woo S, Park J, Lee JY, Kweon IS. CBAM: Convolutional Block Attention Module. In:


1105 Proceedings of the European Conference on Computer Vision. 2018. pp. 3-19.


1106 doi:10.1007/978-3-030-01234-2_1.


1107 &#91;39&#93; Jocher G. Ultralytics YOLOv5 &#91;software&#93;. 2020. doi:10.5281/zenodo.3908559. Available


1108 from: https://github.com/ultralytics/yolov5


1109 &#91;40&#93; Jocher G, Chaurasia A, Qiu J. Ultralytics YOLOv8 &#91;software&#93;. 2023. Available from:


1110 https://github.com/ultralytics/ultralytics


1111 &#91;41&#93; Tian Y, Ye Q, Doermann D. YOLOv12: attention-centric real-time object detectors. arXiv


1112 &#91;Preprint&#93;. 2025. arXiv:2502.12524. doi:10.48550/arXiv.2502.12524.


1113 &#91;42&#93; Jocher G, Qiu J, Liu M, Lyu S, Akyon FC, Kalfaoglu ME. Ultralytics YOLO26: unified real

1114 time end-to-end vision models. arXiv &#91;Preprint&#93;. 2026. arXiv:2606.03748.


1115 doi:10.48550/arXiv.2606.03748.


1116 &#91;43&#93; Liu W, Anguelov D, Erhan D, Szegedy C, Reed S, Fu CY, et al. SSD: Single Shot MultiBox


1117 Detector. In: Computer Vision-ECCV 2016. Cham: Springer; 2016. pp. 21-37.


1118 doi:10.1007/978-3-319-46448-0_2.


1119 &#91;44&#93; Howard A, Sandler M, Chu G, Chen LC, Chen B, Tan M, et al. Searching for MobileNetV3.


1120 In: Proceedings of the IEEE/CVF International Conference on Computer Vision. 2019. pp.


1121 1314-1324. doi:10.1109/ICCV.2019.00140.


1122 &#91;45&#93; Tan M, Pang R, Le QV. EfficientDet: scalable and efficient object detection. In: Proceedings


1123 of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2020. pp.


1124 10781-10790. doi:10.1109/CVPR42600.2020.01079.


1125 &#91;46&#93; Ren S, He K, Girshick R, Sun J. Faster R-CNN: towards real-time object detection with


1126 region proposal networks. IEEE Trans Pattern Anal Mach Intell. 2017;39(6):1137-1149.


1127 doi:10.1109/TPAMI.2016.2577031.


1128 &#91;47&#93; Bai J, Lu F, Zhang K, et al. ONNX: Open Neural Network Exchange &#91;software&#93;. GitHub;


1129 2019. Available from: https://github.com/onnx/onnx


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.17.767224; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.17.767224)

(which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC
105 and is also made available for use under a CC0 license.


1130 &#91;48&#93; ONNX Runtime developers. ONNX Runtime &#91;software&#93;. Available from:


1131 https://onnxruntime.ai/


1132 &#91;49&#93; Burke HR. Studies on the genus Anthonomus in North and Central America (Coleoptera:


1133 Curculionidae) II. The subgenus Anthonomorphus Dietz. Coleopt Bull. 1964;18(1):7-17.


1134 doi:10.5962/p.372349.


1135 &#91;50&#93; Whitehead DR, Chamorro ML, Anderson RS. An illustrated key to the species of Curculio


1136 Linnaeus (Coleoptera: Curculionidae) of North America east of the Mississippi River. Proc


1137 Entomol Soc Wash. 2018;120(3):616-641. doi:10.4289/0013-8797.120.3.616.


1138 &#91;51&#93; Ali F, Qayyum H, Iqbal MJ. Faster-PestNet: a lightweight deep learning framework for crop


1139 pest detection and classification. IEEE Access. 2023;11:104016-104027.


1140 doi:10.1109/ACCESS.2023.3317506.


1141 &#91;52&#93; Dong Q, Sun L, Han T, Cai M, Gao C. PestLite: a novel YOLO-based deep learning


1142 technique for crop pest detection. Agriculture. 2024;14(2):228.


1143 doi:10.3390/agriculture14020228.


1144 &#91;53&#93; Khan A, Malebary SJ, Dang LM, Binzagr F, Song HK, Moon H. AI-enabled crop


1145 management framework for pest detection using visual sensor data. Plants. 2024;13(5):653.


1146 doi:10.3390/plants13050653.


1147 &#91;54&#93; Xiong P, Zhang C, He L, Zhan X, Han Y. Deep learning-based rice pest detection research.


1148 PLoS ONE. 2024;19(11):e0313387. doi:10.1371/journal.pone.0313387.


1149 &#91;55&#93; Liu J, Yu E, Li Y, Zhao Y, Mao B. YOLO-DCPG: a lightweight architecture with dual-channel


1150 pooling gated attention for intensive small-target agricultural pest detection. Front Plant Sci.


1151 2025;16:1716703. doi:10.3389/fpls.2025.1716703.


1152 &#91;56&#93; Joshi A, Guevara D, Earles M. Standardizing and centralizing datasets for efficient training


1153 of agricultural deep learning models. Plant Phenomics. 2023;5:0084.


1154 doi:10.34133/plantphenomics.0084.


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.17.767224; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.17.767224)

(which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC
105 and is also made available for use under a CC0 license.


1155 &#91;57&#93; Sinanc Terzi D. Effect of different weight initialization strategies on transfer learning for


1156 plant disease detection. Plant Pathol. 2024;73:2325-2343. doi:10.1111/ppa.13997.


1157 &#91;58&#93; Vasconcelos C, Birodkar VN, Dumoulin V. Proper reuse of image classification features


1158 improves object detection. In: Proceedings of the IEEE/CVF Conference on Computer


1159 Vision and Pattern Recognition. 2022. pp. 13628-13637.


1160 doi:10.1109/CVPR52688.2022.01326.


1161 &#91;59&#93; Yu C, Xie J, Tony FJA. BGM-YOLO: an accurate and efficient detector for detecting plant


1162 disease. PLoS ONE. 2025;20(5):e0322750. doi:10.1371/journal.pone.0322750.


1163 &#91;60&#93; Song Y, An K, Wang L, Zhou B. Multi-scale closed-loop tuning via spatial frequency


1164 collaborative sensitivity for rice leaf disease detection. PLoS ONE. 2026;21(6):e0351727.


1165 doi:10.1371/journal.pone.0351727.


1166 &#91;61&#93; Karthik RP, Murugesan G, Ballaji HK, Anitha J. UAV-based RGB and multispectral mango


1167 leaf disease detection with benchmarking of YOLOv5 to YOLOv10 and SeqOpt-optimised


1168 YOLOv8 for real-time edge deployment. PLoS ONE. 2026;21(5):e0349855.


1169 doi:10.1371/journal.pone.0349855.


1170 &#91;62&#93; Borianne P, Theveny F, Cabrera-Bosquet L, Lammoglia S-K. CocoaDeep: a preliminary


1171 study of the performance sensitivity to datasets of Faster RCNN, YOLO and transformer


1172 networks for cocoa pod detection. PLoS ONE. 2026;21(7):e0351657.


1173 doi:10.1371/journal.pone.0351657.


1174 &#91;63&#93; Valan M, Makonyi K, Maki A, Vondracek D, Ronquist F. Automated taxonomic identification


1175 of insects with expert-level accuracy using effective feature transfer from convolutional


1176 networks. Syst Biol. 2019;68(6):876-895. doi:10.1093/sysbio/syz014.


1177 &#91;64&#93; Wu X, Zhan C, Lai YK, Cheng MM, Yang J. IP102: a large-scale benchmark dataset for


1178 insect pest recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision


1179 and Pattern Recognition. 2019. doi:10.1109/CVPR.2019.00899.


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.17.767224; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.17.767224)

(which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC
105 and is also made available for use under a CC0 license.


1180 &#91;65&#93; Schneider S, Taylor GW, Kremer SC, Fryxell JM. Getting the bugs out of AI: advancing


1181 ecological research on arthropods through computer vision. Ecol Lett. 2023;26(7):1247

1182 1258. doi:10.1111/ele.14239.


1183 &#91;66&#93; Marais GC, Stratton IC, Johnson AJ, Hulcr J. Progress in developing a bark beetle


1184 identification tool. PLoS ONE. 2025;20(6):e0310716. doi:10.1371/journal.pone.0310716.


1185 &#91;67&#93; Chen J, Kao SH, He H, Zhuo W, Wen S, Lee CH, Chan SHG. Run, don’t walk: chasing


1186 higher FLOPS for faster neural networks. In: Proceedings of the IEEE/CVF Conference on


1187 Computer Vision and Pattern Recognition. 2023. pp. 12021-12031.


1188 doi:10.1109/CVPR52729.2023.01157.


1189 &#91;68&#93; Deng X, Qi L, Liu Z, Liang S, Gong K, Qiu G. Weed target detection at seedling stage in


1190 paddy fields based on YOLOX. PLoS ONE. 2023;18(12):e0294709.


1191 doi:10.1371/journal.pone.0294709.


1192 &#91;69&#93; Lu D, Wang Y. MAR-YOLOv9: a multi-dataset object detection method for agricultural fields


1193 based on YOLOv9. PLoS ONE. 2024;19(10):e0307643. doi:10.1371/journal.pone.0307643.


1194 &#91;70&#93; Qin Y, Liu Q, Liu H, Zhang H, Ji Z, Ganchev I. Lightweight real-time detectors of apple-leaf


1195 diseases operating on embedded devices. PLoS ONE. 2026;21(7):e0352500.


1196 doi:10.1371/journal.pone.0352500.


1197 &#91;71&#93; Jacob B, Kligys S, Chen B, Zhu M, Tang M, Howard A, et al. Quantization and training of


1198 neural networks for efficient integer-arithmetic-only inference. In: Proceedings of the


1199 IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2018. pp. 2704-2713.


1200 doi:10.1109/CVPR.2018.00286.


1201 &#91;72&#93; Wang Z, Ding T, Liang S, Cui H, Gao X. Workpiece surface defect detection based on


1202 YOLOv11 and edge computing. PLoS ONE. 2025;20:e0327546.


1203 doi:10.1371/journal.pone.0327546.


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.17.767224; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.17.767224)

(which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC
105 and is also made available for use under a CC0 license.


1204 &#91;73&#93; Chen Y, Sun H, Tian L, Yang Y, Wang S, Wang T. Detecting infrared UAVs on edge


1205 devices through lightweight instance segmentation. PLoS ONE. 2025;20(8):e0330074.


1206 doi:10.1371/journal.pone.0330074.


1207 &#91;74&#93; Bjerge K, Alison J, Dyrmann M, Frigaard CE, Mann HMR, Hoye TT. Accurate detection and


1208 identification of insects from camera trap images with deep learning. PLOS Sustain


1209 Transform. 2023;2(3):e0000051. doi:10.1371/journal.pstr.0000051.


1210 &#91;75&#93; He K, Gkioxari G, Dollar P, Girshick R. Mask R-CNN. In: Proceedings of the IEEE


1211 International Conference on Computer Vision. 2017. pp. 2961-2969.


1212 doi:10.1109/ICCV.2017.322.


1213 &#91;76&#93; Graving JM, Chae D, Naik H, Li L, Koger B, Costelloe BR, et al. DeepPoseKit, a software


1214 toolkit for fast and robust animal pose estimation using deep learning. eLife. 2019;8:e47994.


1215 doi:10.7554/eLife.47994.


1216 &#91;77&#93; Kirillov A, He K, Girshick R, Rother C, Dollar P. Panoptic segmentation. In: Proceedings of


1217 the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2019. pp. 9404

1218 9413. doi:10.1109/CVPR.2019.00963.


1219 &#91;78&#93; Weyler J, Magistri F, Marks E, Chong YL, Sodano M, Roggiolani G, et al. PhenoBench: a


1220 large dataset and benchmarks for semantic image interpretation in the agricultural domain.


1221 IEEE Trans Pattern Anal Mach Intell. 2024;46(12):9583-9594.


1222 doi:10.1109/TPAMI.2024.3419548.


1223


1224


1225


1226


1227


1228


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.17.767224; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.17.767224)

(which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC
105 and is also made available for use under a CC0 license.


1229 **Supporting Information**


1230 **S1 Fig. Bar plot showing the distribution of labeled images across the training, validation,**


1231 **and test subsets for each of the 14 weevil species.** The x-axis represents the species names,


1232 while the y-axis indicates the number of labeled images per species. Orange bars correspond to


1233 the training set, green bars to the validation set, and blue bars to the test set. A stratified sampling


1234 approach was used to ensure that all subsets retained proportional representation of each


1235 species.


1236 **S2 Fig. Class distribution, spatial distribution, and bounding-box geometry of the**


1237 **annotated RRWIT dataset.** (A) Class frequencies, normalized bounding boxes, object-center


1238 distribution, and width-height distribution. (B) Univariate and pairwise distributions of normalized


1239 x-center, y-center, width, and height. These are dataset-level quality-control diagnostics and do


1240 not independently establish annotation correctness or biological independence among


1241 partitions.


1242 **S3 Fig. Example output from the RRWIT for individual images.** Uploaded image showing


1243 automated detection and species classification with bounding box and predicted label confidence.


1244 **S4 Fig. Validation performance and computational cost of targeted RRWIT architecture**


1245 **modifications.** (A) Seed-42 validation precision, recall, mAP50, and mAP50–95 for the


1246 unmodified RRWIT architecture (A0), P2-to-P3 feature assistance (A1), P3 Convolutional Block


1247 Attention Module refinement (A2), and the combined P2-assistance plus CBAM architecture


1248 (A3). A0 represents the unmodified RRWIT/YOLO11n architecture trained as the seed-42


1249 matched transfer-learning control and is distinct from the historical seed-0 RRWIT.v1.0.0


1250 checkpoint. (B) Parameter count and theoretical computational complexity at an input resolution


1251 of 640 × 640 pixels. Blue circles denote parameters and correspond to the left y-axis; orange


1252 squares denote GFLOPs and correspond to the right y-axis. Lines connect values only to aid


1253 visual comparison among the four discrete architectures. A0 achieved the highest prespecified


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.17.767224; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.17.767224)

(which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC
105 and is also made available for use under a CC0 license.


1254 primary validation metric, mAP50–95 (0.648). A2 incurred negligible additional computational


1255 cost but did not exceed A0, whereas the P2-containing A1 and A3 architectures increased


1256 computational demand without improving mAP50–95. All architectures were screened using the


1257 same seed and validation protocol; the held-out test partition was not used for architecture


1258 selection. Abbreviations: CBAM, Convolutional Block Attention Module; GFLOPs, billions of


1259 floating-point operations; mAP50, mean average precision at intersection over union 0.50;


1260 mAP50-95, mean average precision averaged over intersection-over-union thresholds 0.50 to


1261 0.95.


1262 **S5 Fig. Validation-stage performance of eight object-detection frameworks before locked-**


1263 **test evaluation.** (A) Validation mAP50–95 and (B) validation mAP50 across seeds 42, 123, and


1264 2025. Blue circles represent individual seed runs; open diamonds and horizontal error bars


1265 represent arithmetic mean ± sample SD. (C) Macro and support-weighted F1 calculated at


1266 confidence thresholds selected on validation data. (D) Validation AP stratified by COCO object

1267 size categories for small, medium, and large objects. (E) Per-species validation AP50–95 and


1268 (F) per-species validation AP50 comparing RRWIT (YOLO11n) with YOLO26n; horizontal error


1269 bars represent sample SD across the three seeds, and connecting lines aid within-species


1270 comparison. Overall class-averaged values are reported as mAP, whereas species-specific


1271 values are reported as AP. Abbreviations: AP, average precision; COCO, Common Objects in


1272 Context; F1, harmonic mean of precision and recall; mAP, mean average precision; SD, sample


1273 standard deviation.


1274 **S6 Fig. Validation and locked-test confusion patterns for YOLO12n across three**


1275 **independent training seeds.** Mean row-normalized confusion matrices for YOLO12n on the


1276 validation set (A) and locked test set (B). Each cell represents the arithmetic mean across seeds


1277 42, 123, and 2025 after independent row normalization. Rows are true categories and columns


1278 are predicted categories. Off-diagonal biological cells indicate species substitutions, the


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.17.767224; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.17.767224)

(which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC
105 and is also made available for use under a CC0 license.


1279 Background column represents unmatched ground-truth objects (false negatives), and the


1280 Background row represents unmatched predictions (false positives) by predicted class. The


1281 Background–Background cell is undefined and left blank. Matching and thresholding were


1282 identical to the frozen evaluation procedure used for the other framework-specific confusion


1283 matrices. Abbreviations: F1, harmonic mean of precision and recall; SD, sample standard


1284 deviation.


1285


1286 **S1 Table. Published architecture and implementation characteristics of the eight**


1287 **object-detection frameworks evaluated in this study.** Parameters and GFLOPs are


1288 published specifications for the corresponding reference/pretrained detector variants at the


1289 stated reference input size and are not measurements of the 14-class models fine-tuned in this


1290 study. Software license refers to the implementation used and does not supersede separate


1291 terms that may apply to pretrained weights, training datasets, or downstream deployment.


1292


1293


1294 **S1_Table.** Published architecture and implementation characteristics of the eight object-detection frameworks evaluated in this study.










































|Model|Detector&lt;br&gt;architecture|Architecture&lt;br&gt;reference|Implementatio&lt;br&gt;nused|Reference&lt;br&gt;inputsize&lt;br&gt;(pixels)|Parameters&lt;br&gt;(M)†|GFLOPs†|Software&lt;br&gt;license‡|
|---|---|---|---|---|---|---|---|
|YOLOv5nu|One-stage,&lt;br&gt;anchor-free&lt;br&gt;YOLO|&#91;39&#93;|Ultralytics&lt;br&gt;8.4.118|640|2.6|7.7|AGPL-3.0 /&lt;br&gt;Enterprise|
|YOLOv8n|One-stage,&lt;br&gt;anchor-free&lt;br&gt;YOLO|&#91;40&#93;|Ultralytics&lt;br&gt;8.4.118|640|3.2|8.7|AGPL-3.0 /&lt;br&gt;Enterprise|
|RRWIT&lt;br&gt;(YOLO11n)|One-stage,&lt;br&gt;anchor-free&lt;br&gt;YOLO|&#91;22&#93;|Ultralytics&lt;br&gt;8.4.118|640|2.6|6.5|AGPL-3.0 /&lt;br&gt;Enterprise|
|YOLO12n|One-stage,&lt;br&gt;attention-centric&lt;br&gt;YOLO|&#91;41&#93;|Ultralytics&lt;br&gt;8.4.118|640|2.6|7.6|AGPL-3.0 /&lt;br&gt;Enterprise|
|YOLO26n|One-stage, end-&lt;br&gt;to-end YOLO|&#91;42&#93;|Ultralytics&lt;br&gt;8.4.118|640|2.4|5.4|AGPL-3.0 /&lt;br&gt;Enterprise|
|SSDLite320&lt;br&gt;MobileNetV3-&lt;br&gt;Large|One-stage&lt;br&gt;SSDLite;&lt;br&gt;MobileNetV3-&lt;br&gt;Large backbone|&#91;43,44&#93;|TorchVision&lt;br&gt;0.22.1+cu126|320|3.440|0.58|BSD-3-Clause|
|EfficientDet-D0|One-stage&lt;br&gt;EfficientDet;&lt;br&gt;EfficientNet-B0&lt;br&gt;backbone and&lt;br&gt;BiFPN|&#91;45&#93;|effdet 0.4.1|512|3.88|2.54|Apache-2.0|
|Faster R-CNN&lt;br&gt;MobileNetV3-&lt;br&gt;Large-320 FPN|Two-stage&lt;br&gt;Faster R-CNN;&lt;br&gt;MobileNetV3-&lt;br&gt;Large backbone&lt;br&gt;and FPN|&#91;44,46&#93;|TorchVision&lt;br&gt;0.22.1+cu126|320|19.386|0.72|BSD-3-Clause|



1295 **†** Parameters and GFLOPs are published specifications for the corresponding reference/pretrained detector variants at the stated reference input
1296 size and are not measurements of the 14-class models fine-tuned in this study. For Ultralytics models, these values correspond to the published
1297 fused-model specifications.
1298 **‡** Software license refers to the implementation used in this study and does not supersede separate terms that may apply to pretrained weights,
1299 training datasets, or downstream deployment.


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.17.767224; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.17.767224)

(which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC
105 and is also made available for use under a CC0 license.


1300 **S1 Video. Prototype RRWIT interactive identification and research-assistance workflow.**


1301 The video demonstrates the user-facing RRWIT website, including image upload and live image


1302 capture for automated weevil detection and species identification, together with an exploratory


1303 conversational artificial intelligence interface for information related to RRWIT and the research


1304 program. The conversational component was not systematically evaluated and is not part of the


1305 detection benchmark.


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.17.767224; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.17.767224)

(which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC
105 and is also made available for use under a CC0 license.


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.17.767224; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.17.767224)

(which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC
105 and is also made available for use under a CC0 license.


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.17.767224; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.17.767224)

(which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC
105 and is also made available for use under a CC0 license.


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.17.767224; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.17.767224)

(which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC
105 and is also made available for use under a CC0 license.


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.17.767224; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.17.767224)

(which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC
105 and is also made available for use under a CC0 license.


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.17.767224; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.17.767224)

(which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC
105 and is also made available for use under a CC0 license.


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.17.767224; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.17.767224)

(which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC
105 and is also made available for use under a CC0 license.


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.17.767224; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.17.767224)

(which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC
105 and is also made available for use under a CC0 license.


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.17.767224; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.17.767224)

(which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC
105 and is also made available for use under a CC0 license.


