# Source paper archive

- Source title: From smartphone to gene: Automated image-based phenotyping captures flowering diversity of common buckwheat
- Author(s): Carles Quesada-Traver, Fabian Hess, Flavian Tschurr, Lukas Kronenberg, Gabriele Rimat, Steven Yates, Bruno Studer
- Source landing URL: https://www.biorxiv.org/content/10.64898/2026.11.03.763633v1
- PDF URL: https://www.biorxiv.org/content/10.64898/2026.11.03.763633v1.full.pdf
- PDF SHA-256: 36632df87f486f715b10f8f3dafbde320d7ed854233fac277fc1112469a6e673
- 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.

---

[bioRxiv preprint doi: https://doi.org/10.64898/2026.11.03.763633; this version posted November 4, 2026. The copyright holder for this](https://doi.org/10.64898/2026.11.03.763633)

preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in
[perpetuity. It is made available under a CC-BY 4.0 International license.](http://creativecommons.org/licenses/by/4.0/)


1
# From smartphone to gene: Automated

2
# image-based phenotyping captures

3
# flowering diversity of common buckwheat


4 Carles Quesada-Traver [1†], Fabian Hess [1†], Flavian Tschurr [2], Lukas Kronenberg [3], Gabriele Rimat [4],


5 Steven Yates [1,5] and Bruno Studer [1*]


6 1 Molecular Plant Breeding, Institute of Agricultural Sciences, ETH Zurich, Universitaetstrasse 2,


7 8092 Zurich, Switzerland


8 2 Grassland Sciences, Institute of Agricultural Sciences, ETH Zurich, Universitaetstrasse 2, 8092


9 Zurich, Switzerland


10 3 Precision Phenotyping, Institute of Agricultural and Nutritional Science, Martin Luther University


11 Halle-Wittenberg, Theodor-Lieser-Strasse 11, 06120 Halle (Saale), Germany


12 4 Institute for Crop Science and Plant Breeding, Bavarian State Research Center for Agriculture,


13 LfL, Am Gereuth 2, 85354 Freising, Germany


14 5 Puregene AG, Etzmatt 273, 4314 Zeiningen, Switzerland


15


16 *Corresponding author. Email: bruno.studer@usys.ethz.ch


17 - These authors contributed equally to this work.


18


[bioRxiv preprint doi: https://doi.org/10.64898/2026.11.03.763633; this version posted November 4, 2026. The copyright holder for this](https://doi.org/10.64898/2026.11.03.763633)

preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in
[perpetuity. It is made available under a CC-BY 4.0 International license.](http://creativecommons.org/licenses/by/4.0/)

## 19 Abstract


20 Common buckwheat ( _Fagopyrum esculentum_ Moench) exhibits an indeterminate flowering habit,


21
resulting in the simultaneous occurrence of flowering, seed development and seed ripening. This


22 contributes to yield losses and complicates mechanized harvesting. Therefore, the development


23
of varieties with a desirable flowering pattern is a breeding priority in buckwheat. However,


24
phenotyping of buckwheat flowering under field conditions remains challenging because manual


25 assessments are time consuming and expertise dependent. Here, we show that imaging devices


26
such as smartphones or cameras can be used for phenotyping of flowering traits in a buckwheat


27 collection encompassing 169 accessions. Flowering traits describing flowering time (𝑛=10),


28 flowering intensity (𝑛=6) and abscission (𝑛=6) were extracted from 37,078 images captured over


29
five growing seasons using a Mask R-CNN segmentation model combined with statistical


30 modelling of flowering dynamics. The 22 extracted traits were highly heritable (0.5 < ℎ [2] < 0.93)


31 and correlated well (0.84 < 𝑟< 0.94) with manually assessed flowering stages while providing


32
additional information on flowering intensity. Genome-wide association studies revealed 171


33
genomic regions associated with flowering traits, including marker-trait associations for


34 agronomically important breeding targets. Candidate genes had functional annotations including


35 “Auxin response factor” and “Ethylene-insensitive protein”, gene families involved in hormone


36
signalling and developmental regulation, controlling flowering in other species. These results


37 demonstrate how digital imaging combined with automated feature extraction can be leveraged


38
to gain insight on buckwheat flowering patterns in large scale field trials across different


39
phenotyping platforms. This paves the way to breed for desired flowering patterns using molecular


40 tools in a targeted manner.


41


[bioRxiv preprint doi: https://doi.org/10.64898/2026.11.03.763633; this version posted November 4, 2026. The copyright holder for this](https://doi.org/10.64898/2026.11.03.763633)

preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in
[perpetuity. It is made available under a CC-BY 4.0 International license.](http://creativecommons.org/licenses/by/4.0/)

## 42 1. Introduction


43 Common buckwheat ( _Fagopyrum esculentum_ Moench), hereafter referred to as buckwheat, is a


44 neglected crop having the potential to diversify cereal-dominated farming systems (Joshi et al.,


45 2019). Its resistance to most pests and diseases (Campbell, 1997), its short growing season and


46 the possibility to be managed using machinery make buckwheat a versatile and valuable addition


47
to most crop rotations (Farooq et al., 2016; Strahm et al., 2019). In addition, buckwheat offers a


48 source of gluten-free starch and a favourable protein composition for human consumption


49 (Škrabanja & Kreft, 2016). However, one major issue that limits buckwheat adoption is its


50
prolonged flowering phase (Campbell, 1997).


51
This flowering behaviour complicates mechanical harvesting of buckwheat, as prolonged and


52
asynchronous flowering results in the simultaneous presence of flowers and seeds over an


53 extended period. This disrupts harvest operations and reduces the yield potential (Campbell,


54
1997; Joshi et al., 2019). Early harvest results in the presence of flowers, unripe seeds and high


55 moisture contents in the produce. Late harvest increases the risk of seed scattering and pre

56 harvest sprouting, resulting in yield losses (Funatsuki et al., 2000; Morishita et al., 2020; Ohsawa,


57
2020). In contrast, prolonged flowering is beneficial from an ecological perspective, as it provides


58 nectar and pollen for insects (Cawoy et al., 2009; Strahm et al., 2019). Thus, the lack of buckwheat


59
accessions with flowering patterns tailored to their role in the cropping system hinders increased


60
farmer adoption and the exploitation of its ecological benefits (Hess et al., 2026b).


61
The use of genomic methods to breed for specific flowering traits could accelerate buckwheat


62 improvement, as suggested for other neglected crops (Varshney et al., 2012; Joshi et al., 2019;


63
Yabe and Iwata, 2020; Nature Communications Editorial Staff, 2024). However, using such


64
methods requires efficient phenotyping of buckwheat flowering dynamics. Although BBCH scales


65
have been developed to describe flowering patterns in buckwheat (Arduini et al., 2016; Grimes et


66
al., 2025), their application is laborious and difficult to perform in accessions with high diversity.


67 Stage assignment is not straightforward and requires experience, as developmental stages


68 overlap and vary within plots. As a result, BBCH-based phenotyping does not scale well to large


69
trials. Moreover, BBCH scales do not capture traits related to flowering intensity (i.e. number of


70
flowers or flowering extent). Therefore, efficient and high-throughput approaches are needed to


71 overcome this phenotyping bottleneck (Cobb et al., 2013).


72 In recent years, instance segmentation with deep learning has become a common approach to


73 automate crop phenotyping by detecting and quantifying plant structures directly from images


74 (Murphy et al., 2024). Mask R-CNN, in particular, has been widely adopted because it provides


[bioRxiv preprint doi: https://doi.org/10.64898/2026.11.03.763633; this version posted November 4, 2026. The copyright holder for this](https://doi.org/10.64898/2026.11.03.763633)

preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in
[perpetuity. It is made available under a CC-BY 4.0 International license.](http://creativecommons.org/licenses/by/4.0/)


75 object-level masks that support counting and area-based measurements in complex scenes (He


76 et al., 2017). For example, Mask R-CNN has been used to segment on-tree fruits to enable


77 automated monitoring and size estimation in orchards (Giménez-Gallego et al., 2024). At the


78
other end of the scale, fine-tuned Mask R-CNN models have also been applied to microscopic


79
images for the estimation of stomatal density in apple trees (Zuffa et al., 2025). Image-based


80
modelling approaches have also been used to characterize flowering behaviour. For example,


81
Oppliger et al. (2025) used image-based flower detection to quantify flowering dynamics in peas

82 ( _Pisum sativum_ L.), revealing differences between breeding lines and correlations between flower


83 number and yield. In wheat ( _Triticum aestivum_ L.), Wang et al. (2019) used an image-based


84
classifier to measure the variation in heading date across breeding lines and linked this variation


85
to genomic regions controlling the trait. These examples show that modelling dynamic flowering


86 patterns from image-based features can quantify phenology and link trait variation to genomic


87 regions.


88 In this study, we developed an image-based, automated pipeline to phenotype buckwheat


89
flowering traits relevant for breeding. To ensure that our method can be adapted to different


90 contexts and the equipment available in multi-environment trials we compare the performance


91
of different imaging devices, including smartphones, in different environments. Genome wide


92 association studies are performed on the traits extracted from the image-based pipeline and


93 manual grading of BBCH stage to compare and cross validate the results. In addition, the


94
identified marker-trait associations provide a genetic basis for the targeted breeding of flowering


95 dynamics in buckwheat breeding.

## 96 2. Materials and methods

### 97 2.1. Experimental design, image acquisition and germplasm collection


98 Field trials were conducted in 2023 and 2024 to test the 169 buckwheat accessions of the ETH


99 Zurich/BLW buckwheat collection described in Hess et al. (2026c) at the Plant Research Station


100 of ETH Zurich in Eschikon, Switzerland (47.449 N, 8.682 E, 520 m a.s.l.) and at the Experimental


101 station Baumannshof of the Bavarian State Research Center for Agriculture (LfL) (48.701 N, 8.531


102
E, 360 m a.s.l.). A total of five field trials were conducted including four trials with buckwheat as


103 the main crop (ETH23, ETH24, LfL23, LfL24) and one late-sown trial with buckwheat as the second


104
crop which was located at ETH (ETH24L) (Figure 1A). All field trials followed an augmented row

105 column design, with replicated accessions in complete blocks and spatially distributed check


106
accessions to control within-field heterogeneity. The ETH trials (ETH23 and ETH24) comprised 169


[bioRxiv preprint doi: https://doi.org/10.64898/2026.11.03.763633; this version posted November 4, 2026. The copyright holder for this](https://doi.org/10.64898/2026.11.03.763633)

preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in
[perpetuity. It is made available under a CC-BY 4.0 International license.](http://creativecommons.org/licenses/by/4.0/)


107 accessions, with four of these accessions serving as checks. The LfL trials also comprised 169


108 accessions with seven and one of these accessions serving as checks in 2023 and 2024,


109 respectively. Accessions were replicated three times in both ETH trials, while one and two


110 replications were included in the LfL trials in 2023 and 2024, respectively. A subset of the


111 collection comprising 55 accession was grown in ETH24L in two replicates, with one accession


112 serving as spatial check(Table 1). The image acquisition in Eschikon was performed using a


113 prototype of LITERAL® (Hyphen, Avignon, France). At Baumannshof, images were taken using a


114 Samsung Galaxy S10 smartphone (Samsung Electronics, Suwon, South Korea) in 2023 and an


115 Olympus digital camera model TG-6 (Olympus Corporation, Tokyo, Japan) in 2024. Both devices


116 were attached to a bamboo stick (1.5 m) and the camera height was adjusted according to the


117 maximum canopy height in the plots at the beginning of each imaging campaign. All imaging


118 devices were used with auto settings (focus, ISO, aperture and shutter speed). During dry weather


119 conditions, images were taken twice per week throughout the growing season from germination


120 to harvest. Image acquisition was suspended during rain if it interfered with equipment handling


121 or image quality. Across experiments, plots were imaged between 7 and 35 times (median 9 to


122
31), spanning 65 to 121 days between the first and last image (SD 5.5 to 21.8 days; Table 1).


123 The genetic characterization of the 169 buckwheat accessions based on genome-wide allele


124
frequency fingerprints, as well as the marker filtering procedure applied, were described in Hess


125 et al. (2026c). In short, one hundred leaves of individual plants per accession were sampled in


126 2023 and 2024 to form two pooled samples for each accession. From these samples, DNA was


127 extracted to construct double digest restriction-site associated DNA libraries for sequencing. The


128 sequencing data from both years was pooled (resulting in a single dataset per accession) for


129 subsequent calling of single nucleotide polymorphisms (SNPs) and calculation of allele


130 frequencies representative of the genetic composition of each accession. This resulted in a matrix


131
of 52,182 SNP frequencies after filtering for a minor allele frequency greater than 0.05. Missing


132 values were imputed through linkage-informed k-nearest neighbour-derived weighted mean allele


133 frequency prediction using imputef (v.1.1.1; Paril et al., 2025).

### 134 2.2. Phenology


135 The plant development stage of each plot for the ETH23 and ETH24 trials was assessed twice a


136 week during the growing season. A selection of development stages as described in a proposed


137
BBCH scale for buckwheat was used (Arduini et al., 2016). Specifically, stages 00 (Seed), 10


138 (Cotyledon), 11 (Leaf 1), 12 (Leaf 2), 13 (Leaf 3), 14 (Leaf 4), 50 (Blossoming), 60 (First Flower


139
Open), 62 (Flower Open Terminal Inflorescence, TINF), 70 (Green Fruits), 85 (Brown Fruits), 67 (All


140 Fruits Brown TINF), 86 (End Flowering TINF) of the proposed BBCH scale were assessed. The


[bioRxiv preprint doi: https://doi.org/10.64898/2026.11.03.763633; this version posted November 4, 2026. The copyright holder for this](https://doi.org/10.64898/2026.11.03.763633)

preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in
[perpetuity. It is made available under a CC-BY 4.0 International license.](http://creativecommons.org/licenses/by/4.0/)


141
development stages were numbered (1-13) in the order specified before to account for the

142 temporal sequence of buckwheat development. A monotonically increasing _p_ -spline was fitted to


143 the data of each plot using the _scam_ function of the R package scam (v.1.2.17) with parameters k


144 = 8 and bs = “mpi” to interpolate between recorded stages. Daily BBCH stages were predicted


145 from the splines and used for further analysis. Data of stages 60, 62, 70 and 85 have been


146 published by the authors in a previous preprint manuscript but are presented here again for


147
comparison with image-derived flowering traits (Hess et al., 2026b).

### 148 2.3. Modelling and data analysis


149
Plot-level flowering traits were derived by segmenting flowers in each image and subsequent


150
modelling over time to quantify flowering dynamics for downstream analyses (Figure 1). All


151 calculations were performed in R (v.4.4.0) unless stated otherwise.


152
#### 2.3.1. Image processing and flower segmentation


153
Each original image was pre-processed by applying a fixed border crop of 15% to exclude non-plot


154 regions from the picture. The cropped images were then divided into four quadrants (see Figure


155
1B). A set of 241 quadrants from randomly chosen images taken in the ETH23 field trial spanning


156
the growing season was selected, and flower clusters were manually annotated using the polygon


157 tool on the VGG Image Annotator (VIA) v.3.0.11 (Dutta & Zisserman, 2019). This set was randomly


158 divided into two groups, which constituted the training (𝑛 = 192) and the validation (𝑛 = 49) sets,


159 respectively. Mask R-CNN with a ResNet-50 backbone and a Feature Pyramid Network (FPN)


160
initialized with pre-trained COCO weights (Lin et al., 2014) was used for flower instance


161 segmentation. The model was trained for a total of 1,200 epochs using stochastic gradient


162 descent with momentum (µ = 0 _._ 9). The loss function to be minimized, combined classification


163 and bounding box regression losses for both the Region Proposal Network (RPN) and the Fast R

164 CNN head, along with a mask prediction loss for the Mask R-CNN head, regularized by L2 weight


165 decay (𝜆 = 1 × 10 _[−]_ [4] ), as described in Faster R-CNN (Ren et al., 2017) and Mask R-CNN (He et al.,


166 2017).


167 The base learning rate was set to 0.01 with a short linear warm-up and subsequent step decays.


168
An effective batch size of approximately eight was achieved through gradient accumulation. In


169 addition, model checkpoints were saved every 50 epochs and used to calculate metrics on the


170 validation set. Training was performed using TensorFlow (v.1.15.5) and TensorPack (v.0.11.0) in


171 Python (v.3.8.13) running on a single NVIDIA GeForce RTX 3080 Ti (NVIDIA Corporation, Santa


172 Clara, CA, USA). The best checkpoint, based on training and validation metrics, was selected and


173 retained for further predictions.


[bioRxiv preprint doi: https://doi.org/10.64898/2026.11.03.763633; this version posted November 4, 2026. The copyright holder for this](https://doi.org/10.64898/2026.11.03.763633)

preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in
[perpetuity. It is made available under a CC-BY 4.0 International license.](http://creativecommons.org/licenses/by/4.0/)


174
Once selected, the model was used to generate the prediction masks of the flower clusters in all


175 the collected images. For each image, predictions were produced on the cropped image and on


176
each of the four quadrants to improve detection across different scales. Quadrant outputs were


177 mapped back onto the cropped image, and the predictions were combined into a single collection


178 of polygon masks per image. Any overlapping or duplicated masks were merged into single masks


179
to yield one final set of masks per image. Finally, for each image the percentage of pixels inside


180
the masks relative to the total number of pixels was calculated and used as a proxy for flowering


181 intensity in subsequent analyses and is hereafter referred to as Flower %. Image and mask


182 manipulation were performed using functions from the Python packages cv2 (v.4.6.0), shapely


183 (v.9.2.0) and PIL (v.2.9.6).

#### 184 2.3.2. Flowering dynamics modelling and parameter extraction


185
To extract flowering-related phenotypes at the plot level, the observations of each time series


186
were imputed to daily resolution using cubic smoothing splines, fitted with the smooth.spline


187 function from the stats package (v.4.5.1) in R. The measurement period was then divided into two


188
phases: flowering and abscission. The flowering phase was defined by the start of the


189
measurements and the time point of maximum flower coverage, while the abscission phase


190 extended from this point to the end of the measurement period. A diagram showing how this


191
sequential model is fitted, along with all the extracted parameters, is illustrated in Figure 2 and a


192 recollection of all the traits, with information on their calculation and type can be seen in Table


193
S1. All date-related parameters were normalized to days after sowing (DAS). The flowering phase


194 was modelled using a three-parameter self-starting logistic function described in Equation 1:


𝐴
195 𝑓 𝑓𝑙𝑜𝑤𝑒𝑟𝑖𝑛𝑔 (𝑡) = 1+𝑒 [ (𝑡50−𝑡)] ⁄ [𝑠] (1)


196 where 𝑓 𝑓𝑙𝑜𝑤𝑒𝑟𝑖𝑛𝑔 (𝑡) represents the Flower % at time 𝑡 (days), 𝐴 is the asymptote maximum flower


197 coverage (%), 𝑡 50 denotes the inflection point corresponding to mid flowering (days) and 𝑠 is a


198
scaling parameter controlling the steepness of the curve (days). The model was fitted by nonlinear


199 least squares using the nls function with the SSlogis formulation from the R package stats


200 (v.4.5.1). From each fitted curve, the coefficients 𝐴, 𝑡 50 and 𝑠 were retrieved as Asymptote (%),


201 Mid Flowering (days) and Flowering Scale (days). Based on the sensitivity of the image


202 segmentation model, Start Flowering DAS (days) was defined as the first date when the


203
percentage of flowers covered at least 2% of the image area. The first derivative of the fitted curve,


204 𝑑𝑓 𝑓𝑙𝑜𝑤𝑒𝑟𝑖𝑛𝑔 (𝑡)/𝑑𝑡, was used to identify the time point at which the instantaneous rate of increase


205
in flower coverage was maximal, this time point was defined as Max Flowering Rate DAS (days),


206
and the flower coverage at this time point was recorded as Max Flowering Rate (%). Following an


[bioRxiv preprint doi: https://doi.org/10.64898/2026.11.03.763633; this version posted November 4, 2026. The copyright holder for this](https://doi.org/10.64898/2026.11.03.763633)

preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in
[perpetuity. It is made available under a CC-BY 4.0 International license.](http://creativecommons.org/licenses/by/4.0/)


207 approach similar to that used to characterize senescence dynamics in wheat (Tschurr et al. 2024),


208 the second derivative of the fitted logistic function, 𝑑 [2] 𝑓 𝑓𝑙𝑜𝑤𝑒𝑟𝑖𝑛𝑔 (𝑡)/𝑑𝑡 [2], was used to determine


209 End Flowering DAS (days) and Max Flowering (%), corresponding to the point at which the


210
acceleration of flowering declines towards the stationary phase of the curve. Finally, the area


211
under the fitted logistic curve (Flowering AUC, % days) was calculated by numerical integration of


212
the fitted logistic function over the flowering phase. The abscission phase was modelled using a


213 linear function of time as described in Equation 2:


214 𝑓 𝑎𝑏𝑠𝑐𝑖𝑠𝑠𝑖𝑜𝑛 (𝑡) = 𝛽 0 + 𝛽 1 𝑡 (2)


215 where 𝑓 𝑎𝑏𝑠𝑐𝑖𝑠𝑠𝑖𝑜𝑛 (𝑡) represents the Flower % at time 𝑡 (days). The model was fitted with ordinary

216 least squares using the lm function from base R. The estimated coefficient 𝛽 1 was interpreted as


217 Abscission Rate (% days [-1] ), while 𝛽 0 represented the Abscission Intercept Amplitude (%). The


218 point at which the fitted line crossed the time axis (𝑓 𝑎𝑏𝑠𝑐𝑖𝑠𝑠𝑖𝑜𝑛 (𝑡) = 0) during the abscission phase


219 was recorded as Abscission End (days), from which a normalized value (Abscission Projected End


220 DAS (days)) was also calculated.


221
In addition to the parameters directly obtained from the flowering and abscission models, six


222 duration parameters and three amplitude parameters were calculated. Flower Duration (days)


223 was described as the interval from Start Flowering DAS to End Flowering DAS. This duration was


224 further divided into Flower Duration Early (days), spanning from Start Flowering DAS to Max


225 Flowering Rate DAS, and Flower Duration Late (days), from Max Flowering Rate DAS until End


226
Flowering DAS. Abscission Duration (days) was defined as the interval between End Flowering


227 DAS and Abscission Projected End DAS. Extending this timeframe, the Flower to Maturity Period


228
(days) was defined as the interval between Start Flowering DAS and Abscission Projected End


229
DAS. While Late Flowering to Abscission End (days) was defined as the period between Max


230
Flowering Rate DAS and Abscission Projected End DAS. Changes in flower coverage during the


231
flowering phase were quantified using three amplitude parameters. Flower Amplitude (%) was


232
calculated as the difference between Max Flowering and Start Flowering (2%). This total change


233
was further divided into Flower Amplitude Early (%), defined as the difference in flower coverage


234
between Max Flowering Rate and Start Flowering (2%), and Flower Amplitude Late (%), defined as


235
the difference between Max Flowering and Max Flowering Rate.


236 To assess the goodness-of-fit, the 𝑅 [2] statistic was calculated from the sum of the squared errors


237
for the fit of the flowering and abscission phases separately. Parameters extracted from these fits


238 with a 𝑅 [2] lower than 0.5 were excluded from the subsequent analysis.


[bioRxiv preprint doi: https://doi.org/10.64898/2026.11.03.763633; this version posted November 4, 2026. The copyright holder for this](https://doi.org/10.64898/2026.11.03.763633)

preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in
[perpetuity. It is made available under a CC-BY 4.0 International license.](http://creativecommons.org/licenses/by/4.0/)

#### 239 2.3.3. Spatial correction and repeatability


240
The traits extracted at the plot level by the flowering dynamics model were further analysed using


241 a two-stage linear mixed model as described in (Roth et al., 2021) and (Roth et al., 2024). In the


242
first stage, adjusted genotype means of all traits were calculated for each year-site using the R

243 package SpATS (v.1.0.19) (Rodríguez-Álvarez et al., 2018). The model parameters were specified


244 as:


245 𝑌 𝑖𝑗𝑘 = 𝑓 (𝑟, 𝑐) + 𝐺 𝑖 + 𝑅 𝑗 + 𝐶 𝑘 + 𝜀 𝑖𝑗𝑘 (3)


246 where 𝑌 𝑖𝑗𝑘 is the parameter value of the plot in the 𝑖 [𝑡ℎ] genotype, 𝑗 [𝑡ℎ] year-site and of the


247 𝑘 [𝑡ℎ] replication, 𝑓 (𝑟, 𝑐) is a smooth bivariate surface defined over rows and ranges, 𝐺 𝑖 is the effect


248 of the 𝑖 [𝑡ℎ] genotype while 𝑅 𝑗 and 𝐶 𝑘 are the effects of the 𝑗 [𝑡ℎ] row and 𝑘 [𝑡ℎ] range, respectively. The


249 plot residuals are denoted as 𝜀 𝑖𝑗𝑘 . In Equation 3, 𝐺 𝑖 was set as random to calculate repeatability


250 (Oakey et al., 2006), while for the calculation of best linear unbiased estimators (BLUEs) the term


251
was set as fixed. Moreover, the diagonal of the inverse of the variance-covariance matrix from this


252 model was extracted and used as weights in the next stage.

#### 253 2.3.4. Genotype means across year-sites and heritability


254 To calculate genotype means across year-sites, the R package ASReml (v.4.1.0.176) was used to


255 specify the model:


256 𝑌 𝑖𝑗 = µ + 𝐺 𝑖 + 𝑌 𝑗 + 𝐺𝑌 𝑖𝑗 (4)


257 where 𝑌 𝑖𝑗 are the BLUEs calculated in the previous stage, µ the global intersect, 𝐺 𝑖 the genotype

258 effect, 𝑌 𝑗 the year-site effect and 𝐺𝑌 𝑖𝑗 represents the weighted genotype by year-site interaction


259 modelled through a diagonal error variance structure for genotypes nested within year-sites.


260 Since we used adjusted genotype means per year-site, the GxE interaction is confounded with the


261 residual error and no explicit quantification of this variance component can be given. The 𝐺 𝑖 term


262 in Equation 4 was set as random to calculate BLUPs and heritabilities (Cullis et al., 2006) while it


263
was specified as fixed for the calculation of BLUEs across years.

#### 264 2.3.5. Correlations between phenotypic values


265
Pairwise Pearson correlation coefficients were calculated at two different levels of the phenotypic


266 data using the cor function in R with use=“pairwise.complete.obs”. First, using the BLUEs per year,


267 correlations were calculated for all combinations of the available image-derived datasets, namely


268
ETH24, ETH23, LfL23, LfL24 and ETH24L. Second, focusing on the field trials ETH23 and ETH24,


269
the BLUPs resulting from Equation 4 were correlated in two different ways: within image-derived


[bioRxiv preprint doi: https://doi.org/10.64898/2026.11.03.763633; this version posted November 4, 2026. The copyright holder for this](https://doi.org/10.64898/2026.11.03.763633)

preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in
[perpetuity. It is made available under a CC-BY 4.0 International license.](http://creativecommons.org/licenses/by/4.0/)


270 traits and between image-derived and manually measured traits. For this, two-sided _p_ -values


271 were calculated for each comparison and adjusted for multiple testing using the Benjamini–


272 Hochberg false discovery rate at three different significance levels: 𝛼 < 0 _._ 05, 𝛼 < 0 _._ 01 and 𝛼 <


273 0 _._ 001. Based on these values, the function hclust from stats (v.4.5.1) was used to perform


274 hierarchical clustering to establish groups of similar traits for subsequent analysis.

#### 275 2.3.6. Principal component analysis


276 The accessions of the buckwheat collection were previously grouped into seven subclusters using


277 NGSadmix as described in Hess et al. (2026c). In addition, the accessions were also labelled


278
according to the morphology of their flowers (which reflects self-compatibility) and their terminal


279
flower cluster. A principal component analysis (PCA) was performed on the centred and scaled


280
BLUEs calculated over ETH23 and ETH24 for all 22 image-derived flowering traits using the

281 function prcomp of the R package stats (v.4.5.1) and was used to assess differences in flowering


282
behaviour between different subclusters and the previously determined flowering types.

#### 283 2.3.7. Genome-wide association studies for image-derived and manually


284
#### measured flowering traits


285 Genome-wide association studies (GWAS) were performed using GAPIT (v.3.5.0) in R (Wang and


286 Zhang, 2021) on image-derived and manually measured datasets from the ETH trials (2023, 2024


287 and combined). For each trait-dataset combination, four methods implemented in GAPIT were


288 tested: BLINK (Huang et al., 2019), FarmCPU (X. Liu et al., 2016), mixed linear model (MLM) (Yu et


289 al., 2006) and multi-locus mixed model (MLMM) (Segura et al., 2012). The phenotypes were year

290 site BLUEs, and the genotypes were supplied as an imputed allele-frequency matrix comprising


291
52,182 SNPs. The structure of the population was controlled using the first three principal


292 components of the genotype matrix and relatedness was modelled using the GAPIT


293 implementation of the genomic kinship matrix proposed by (VanRaden, 2007). For each run,


294 marker-trait association significance was assessed using the Benjamini-Hochberg FDR at 𝛼 <


295 0 _._ 05.


296 Two complementary approaches were used to summarize and visualize patterns of marker-trait


297 associations from the results of 420 GWAS analyses. First, exact SNP overlaps among previously


298
defined trait groups were visualized by generating an UpSet plot (Lex et al., 2014) from all


299 significant marker-trait associations. Second, 𝑝-values of significant marker-trait associations


300 involving the same marker across models and datasets were combined by Fisher’s combined


301 probability test (Fisher, 1992) to produce a single combined 𝑝-value per position. The combined


[bioRxiv preprint doi: https://doi.org/10.64898/2026.11.03.763633; this version posted November 4, 2026. The copyright holder for this](https://doi.org/10.64898/2026.11.03.763633)

preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in
[perpetuity. It is made available under a CC-BY 4.0 International license.](http://creativecommons.org/licenses/by/4.0/)


302 −𝑙𝑜𝑔 10 (𝑝) were plotted at genomic coordinates to visualize the distribution and significance of


303 grouped marker-trait associations.


304 Since exact SNP overlap may miss associations originating from the same genetic signal, linkage


305 disequilibrium (LD)-supported concordance between image-derived and manually measured


306 GWAS results was assessed. Pairwise LD was calculated from the imputed allele-frequency


307 matrix as dosage-based 𝑟 [2], defined as the squared Pearson correlation coefficient between SNP


308
vectors across accessions (Hill & Robertson, 1968). Background-adjusted LD decay was first


309 estimated separately for each chromosome using all SNP pairs within 2 Mb, grouped into 5 kb


310 distance bins containing at least 100 pairs. Initial LD was estimated from pairs separated by 5 to


311 50 kb, whereas background LD was estimated from pairs separated by 500 kb to 2 Mb. A


312
chromosome-specific LD window was defined as the first distance at which the LOESS-smoothed


313 median 𝑟 [2] crossed the background-adjusted half-decay threshold for three consecutive bins,


314
with a minimum window of 50 kb. These chromosome-specific LD windows (Table S2) were then


315 used to construct LD-supported blocks by connecting SNP pairs within the corresponding


316 window when 𝑟 [2] ≥0.20. Connected components were used as LD blocks, with their maximum


317
span capped at the chromosome-specific LD window. Image-derived and manually measured


318 associations were considered concordant when they involved the same SNP or when distinct


319
SNPs occurred within the same LD block, and were classified as flowering, abscission or mixed


320 trait-category matches.


321 To further investigate the associations of the image-derived phenotypes, all genes directly co

322
localizing with markers involved in significant marker–trait associations were retrieved. For


323
selected associations, regional GWAS profiles were inspected across datasets and models, and


324 local LD was visualized using the focal SNP together with 12 neighbouring SNPs on each side.


325 Pairwise LD in these local plots was calculated from the allele-frequency matrix as described


326 above. The predicted protein sequences of these genes were extracted from the genome


327 assembly Tuka (Hess et al., 2026a) and used as input for EggNOG-Mapper (v.2.1.13;


328 Cantalapiedra et al., 2021) with default parameters. This tool uses precomputed orthologous


329 groups and phylogenies from the EggNOG (v.5.0.0) database (Huerta-Cepas et al., 2019) to


330
transfer functional information to the queried sequences. Additionally, for all significant marker

331 𝑅 [2]
trait associations within a gene, a linear model was fitted to model the allelic effect, where the


332
coefficient was used to describe the percentage of phenotypic variation explained.


[bioRxiv preprint doi: https://doi.org/10.64898/2026.11.03.763633; this version posted November 4, 2026. The copyright holder for this](https://doi.org/10.64898/2026.11.03.763633)

preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in
[perpetuity. It is made available under a CC-BY 4.0 International license.](http://creativecommons.org/licenses/by/4.0/)

## 333 3. Results


334
### 3.1. Image segmentation model for flower percentage estimation


335 During the image segmentation model training, both box and mask losses initially decreased until


336 a steady state was reached around epoch 250 (Figure 3A). This state was maintained until the end


337 of training. A similar pattern appeared in validation losses computed at 50-epoch intervals from


338 saved checkpoints. The correlations between the predicted areas and the ground-truth mirrored


339 the stability of the loss dynamics (Figure 3B), where all checkpoints showed very high correlations


340 for both sets (𝑟 > 0 _._ 97). The unprocessed predicted masks from the training set showed a higher


341 correlation with the ground truth compared to the validation set. After merging overlapping


342 masks, all the correlations increased to 𝑟 > 0 _._ 99, indicating that the raw output occasionally


343 predicted the same mask multiple times. The best-performing checkpoint based on correlation


344 (after merging the masks) occurred at epoch 550 and is shown in detail in Figure 3C. The


345 scatterplot for this checkpoint shows a slight underprediction of Flower %, visible as a small shift


346 of the points from the 1:1 line. This bias is negligible for downstream analyses because it is


347
systematic across images processed by the same model and does not affect the overall


348
distribution. Images without annotated flowers showed an average predicted flower area of 0.5%,


349 with some up to 3% (Figure 3D). Based on this, a percentage of 2% was set as the detection limit


350 for our trained image segmentation model.


351
### 3.2. Modelling of buckwheat flowering dynamics


352 Compared to the other experiments, the ETH23 and ETH24 trials were characterized by a higher


353 data density, with a median of 22 and 31 images per plot and time-series durations of 95 and 121


354 days on average for 2023 and 2024, respectively (Table 1). All the datasets showed an overall


355 𝑅 [2]
satisfactory goodness-of-fit, with median values above 0.85 for both the flowering and


356 abscission phases across all experiments (Figure 4A). The three-parameter logistic function


357
describing the flowering phase achieved higher goodness-of-fit than the linear function used for


358 the abscission phase, with median 𝑅 [2] values ranging from 0.98 to 0.99 and from 0.85 to 0.91,


359 respectively. After filtering for fits with 𝑅 [2] - 0 _._ 5 three plots were excluded for the flowering phase


360 and 35 plots for the abscission phase, and their corresponding parameters were discarded (Table


361 1). Notably, the ETH23 trial showed the highest number of discarded abscission-phase models,


362
with 35 removed out of 594. An example of model fits for a single accession (FE 268) across all


363 datasets is shown in Figure 4B, illustrating some of the contrasting metrics presented in Table 1,


364 such as plots per accession, number of time points and mean duration.


[bioRxiv preprint doi: https://doi.org/10.64898/2026.11.03.763633; this version posted November 4, 2026. The copyright holder for this](https://doi.org/10.64898/2026.11.03.763633)

preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in
[perpetuity. It is made available under a CC-BY 4.0 International license.](http://creativecommons.org/licenses/by/4.0/)


365
### 3.3. Variation in flowering traits across sowing time, accessions and 366 environments


367
Buckwheat grown as a second crop at ETH (ETH24L) reached all flowering phases earlier (days


368 after sowing) than buckwheat grown as a main crop (combined ETH23 and ETH24; Figure 5A1).


369 Flowering intensity and Flowering AUC were also lower in the second-crop trial (Figures 5A2–A3),


370 and all differences were statistically significant (Wilcoxon-test: 𝑝 _<_ 0.001). Moreover, the variation


371 between accessions of buckwheat as a second crop was lower for time point traits (Figure 5A1)


372 and Flowering AUC (Figure 5A3) while this pattern was not observed for flowering intensity traits.


373 For example, on average buckwheat sown as the main crop reached Projected End of Abscission


374 DAS after 124 days, and the trait showed a standard deviation of 15.86 days, while the buckwheat


375 sown after wheat reached the same development stage already after 79 days, and the standard


376
deviation among accessions was only 7.03 days. These patterns as well as the significant


377
differences between buckwheat grown as a main crop and as a second crop remained persistent


378 when only overlapping accessions between the trials were considered.


379 The buckwheat collection analysed in this study is composed of several subclusters (Hess et al.,


380 2026b) and contains a total of nine self-compatible and eleven accessions with determinate


381
growth habit. These different subclusters showed distinct flowering dynamics as shown in the


382 PCA (based on combined data from ETH23 and ETH24) in Figure 5B. Accessions from subclusters


383 1 (mainly originating from East Asia) and 7 (diverse origin) were associated with late Start


384 Flowering DAS and Projected End of Abscission DAS while subcluster 3, which includes


385
accessions from the Himalayan region, showed moderately late flowering and slow Abscission


386 Rate (higher, i.e., less negative values for abscission slope indicate a slower decline during


387 abscission phase). Subclusters 2 and 4, representing Central and Eastern accessions,


388 respectively, showed early and short generative development. Determinate and self-compatible


389
accessions (from subcluster 5) both showed low flowering intensity and low Flowering AUC.


390
Comparison of pairwise correlation coefficients for flowering traits observed in different


391 environments shows that some traits are highly correlated in all comparisons, while others have


392
consistently low correlations (Figure 5C). Traits related to time points such as Mid flowering DAS,


393 Max Flowering Rate DAS, Start Flowering DAS or Projected End of Abscission DAS showed the


394 highest average correlation over all year-site comparisons (0 _._ 64 < 𝑟 < 0.77). Correlations


395 between year-sites were lower (𝑟 < 0 _._ 6) for traits describing durations of flowering phases or


396 flowering amplitudes. Low correlations (𝑟 < 0 _._ 4) were also recorded for Abscission Rate and


397
Flowering AUC. Comparing the average correlation coefficients between the year-site


[bioRxiv preprint doi: https://doi.org/10.64898/2026.11.03.763633; this version posted November 4, 2026. The copyright holder for this](https://doi.org/10.64898/2026.11.03.763633)

preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in
[perpetuity. It is made available under a CC-BY 4.0 International license.](http://creativecommons.org/licenses/by/4.0/)


398 comparisons revealed that correlations between the trial with buckwheat as a second crop with


399 trials where buckwheat was planted as a main crop were weaker (𝑟̅ < 0 _._ 5) compared to


400 correlations between main crop trials (𝑟̅ > 0 _._ 5, except for the comparison LfL23 and LfL24). The


401
highest average correlations coefficients were found for the comparison ETH23 and ETH24 with


402 𝑟̅
equal to 0.62. However, even for comparisons of year-sites with different sowing times,


403 correlations for time point traits remained high with values above 0.6 (𝑟). The same pattern as for


404 the correlations between year-sites was observed in broad-sense heritabilities, i.e., heritability


405 was higher for traits related to time points compared to traits describing the durations of


406
phenological phases or flowering amplitudes (Tables S3-S5).


407
### 3.4. Comparison of image-derived and manually measured flowering 408 traits


409 Given the availability of BBCH traits scored manually for the ETH23 and ETH24 trials (Table S6),


410
together with the high temporal resolution used to extract image-derived flowering phenotypes,


411 these datasets were used for subsequent comparisons between both approaches. The pairwise


412 Pearson correlations of the BLUPs over years between image-derived traits presented high values


413 of 𝑟 within themselves and broadly fell into three major groups (Figure 6A). The phenotypes


414 describing the Flower % (flowering 𝑦-axis) highly correlated among themselves, with values


415 ranging from 0.77 to 1 (mean of 0.92), and constituted a clear group named hereafter Image

416
derived Flower Value. The group describing time points as well as periods during the flowering


417 phase (flowering 𝑥-axis) correlated with high to modest 𝑟 values with an average (𝑟̅) of 0 _._ 69 in a


418 group named Image-derived Flower Period. The last group of phenotypes, Image-derived


419 Abscission, was constituted by most of the parameters extracted from the abscission modelling


420 phase that also presented high 𝑟 values with 𝑟̅ = 0 _._ 78. It is important to note that the traits


421 Abscission Intercept Amplitude and Flowering AUC could have been assigned to other groups


422 given their meaning and 𝑟 values but were assigned to the Image-derived Flower Period group for


423 their overall higher 𝑟 with the rest of the traits in this group.


424 When comparing image-derived phenotypes with temporally ordered BBCH stages, additional


425
patterns were revealed (Figure 6B). The first image-derived group (Image-derived Flower Value)


426 showed a consistent correlation pattern with the manually measured traits, with a weak but


427
significant negative correlation with values ranging from -0.17 to -0.50 with the stages of


428
vegetative development, and a weak but significant correlation ranging from 0.51 to 0.6 with End


429
Flowering TINF, which describes the end of flowering in the terminal inflorescence. All correlations


430 between manually measured traits and image-derived traits in this group were below |𝑟| = 0 _._ 6.


[bioRxiv preprint doi: https://doi.org/10.64898/2026.11.03.763633; this version posted November 4, 2026. The copyright holder for this](https://doi.org/10.64898/2026.11.03.763633)

preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in
[perpetuity. It is made available under a CC-BY 4.0 International license.](http://creativecommons.org/licenses/by/4.0/)


431 The second image-derived group (Image-derived Flower Period) also presented a consistent


432
correlation pattern with the manually measured BBCH stages during flowering development


433 (Manual Flowering). Timepoint-related phenotypes, such as Start Flowering DAS, Max Flowering


434 Rate DAS, and Mid Flowering showed the highest correlations with up to 𝑟= 0.92, whereas


435 period-related phenotypes presented values around 0.75. Finally, the third group of image

436 derived phenotypes related to abscission (Image-derived Abscission) also showed a consistent


437 pattern of correlations with the manually measured BBCH stages. These traits showed negative


438 correlations with early BBCH stages starting at 𝑟= -0.52, which increased along the subsequent


439 stages with some values up to 0.69. Highest correlations were found with the last measured BBCH


440 stages (Manual Abscission) with values ranging from 0.50 to 0.94. It has to be noted that


441 Abscission Rate, although following the same pattern, presented an overall weaker correlation


442 with all the BBCH stages compared to the other phenotypes from the Image-derived Abscission


443 group.

### 444 3.5. Comparison of genome-wide association studies between image- 445 derived and manually measured traits


446 Image-derived phenotyping methods generated a higher number of associations (𝑛 = 239)


447 compared to manual scoring methods (𝑛 = 170) across the ETH23, ETH24, and combined


448 datasets when comparing Flower Period and Abscission traits (Figure 7A). Most of these


449 associations originated from the Image-derived Flower Period group (𝑛 = 197), which is the one


450 constituted by the highest number of traits (𝑛 𝑡 = 10). Within these groups, only three significant


451 associations were shared between Image-derived Flower Period and Image-derived Abscission.


452 For BBCH-based phenotyping scores, a similar trend was observed, with Manual Flowering as the


453 group with the most associations (𝑛 = 118), nine of which were shared with the group Manual


454 Abscission. Generally, 26 associations involved the same markers between the phenotyping


455 methods, with 16 of these markers present in both Image-derived Flower Period and Manual


456
Flowering. A general overview of the 463 significant marker-trait associations for both phenotyping


457 methods in their chromosomal context is displayed in Figure 7B. The 26 direct common


458 associations were found in six of the eight chromosomes (Chr). Chr3, Chr7 and Chr8 harbour the


459 highest number of associations (𝑛 = 21). Accounting for LD between nearby markers 21


460 additional coincidences between image-derived and manually measured GWAS associations


461
were identified. These coincidences mostly clustered around loci already detected through exact


462
SNP overlap, rather than defining new signals. Only four pointed to additional shared regions


463 without a nearby direct overlap, located on Chr1, Chr2 and Chr6. Additional regions without direct


464 or LD-supported coincidence but with close associations between methods can also be


[bioRxiv preprint doi: https://doi.org/10.64898/2026.11.03.763633; this version posted November 4, 2026. The copyright holder for this](https://doi.org/10.64898/2026.11.03.763633)

preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in
[perpetuity. It is made available under a CC-BY 4.0 International license.](http://creativecommons.org/licenses/by/4.0/)


465
observed. Some interesting regions fulfilling these criteria are the first megabase pairs from Chr6,


466 the last from Chr1, or around the position 50Mb on Chr4. Figures S1–S2 report the complete set


467 of marker-trait associations for each trait across datasets and GWAS models.


468
### 3.6. Putative candidate genes governing flowering dynamics in 469 buckwheat


470 The genetic context of the Image-derived phenotype associations was examined in detail using


471
the uniquely identified 343 marker-trait associations. Of these, 265 were found to co-localize with


472 171 annotated genes, 93 of which contained markers located within the putative coding sequence


473 (CDS; Table 2). Four gene-overlapping marker-trait associations were selected for detailed


474
inspection using regional GWAS profiles, local LD heatmaps and marker-effect plots (Figure 8).


475
Across the four loci, the regional GWAS profiles showed localized association peaks at the focal


476 markers. The local LD heatmaps showed neighbouring markers with variable pairwise 𝑟 [2], but did


477 not show continuous high-LD blocks spanning the full plotted intervals. The marker


478
“chr1_163426249_G/A” affecting Mid Flowering, co-localizes with the gene “evm.TU.chr1.4316”,


479
annotated as an “Auxin response factor (ARF)” (Figure 8A). This marker had a strong and significant


480 effect across all three datasets, with 𝑅 [2] values of 0.374, 0.351 and 0.397, indicating that a


481 substantial proportion of the trait variance was explained by this single marker. The marker


482
“chr8_79940130_C/T” affecting Abscission Rate, co-localizes with the gene “evm.TU.chr8.992”


483 annotated as “Ethylene-insensitive protein” (Figure 8B). The marker “chr6_16913820_G/C”


484
affecting Flowering Amplitude, is located within the gene “evm.TU.chr6.1780” annotated as a


485
“Nuclear chaperone” (Figure 8C). Finally, the marker “chr7_633641_G/A” affecting Start Flowering


486 DAS, co-localizes with the gene “evm.TU.chr7.98” annotated as a “Diphthamide biosynthesis


487 𝑅 [2] values
protein” (Figure 8D). The latter three markers showed weaker effects in ETH23, with


488 ranging from 0.048 to 0.199, when compared to the ETH24 and combined datasets, with 𝑅 [2] values


489 between 0.213 and 0.322. Further details showing all marker-trait associations that fell within


490 annotated genes across the three datasets, are provided in Table S7.

## 491 4. Discussion


492 Our image-based segmentation approach yielded accurate predictions of Flower %, with a strong


493 correlation between predicted and manually measured values (𝑟 > 0 _._ 99). The robustness of the


494
image-derived flowering traits describing buckwheat flowering patterns was further supported by


495 high repeatability and heritability for most traits. In addition, the image-derived traits accurately


496 predict all manually measured BBCH stages related to flowering (𝑟 > 0 _._ 9) and seed formation


[bioRxiv preprint doi: https://doi.org/10.64898/2026.11.03.763633; this version posted November 4, 2026. The copyright holder for this](https://doi.org/10.64898/2026.11.03.763633)

preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in
[perpetuity. It is made available under a CC-BY 4.0 International license.](http://creativecommons.org/licenses/by/4.0/)


497 (𝑟 > 0 _._ 85), demonstrating their biological and agronomic relevance. Most importantly, our


498
phenotyping approach encompassed different sowing times (main and second crop),


499 environmental conditions (sites and years), imaging devices (LITERAL®, smartphone, high


500
resolution digital camera) and field staff responsible for image acquisition. This demonstrates


501
that the workflow described here (image-acquisition, image-segmentation, quantification of


502
flowering dynamics) represents a robust and broadly applicable phenotyping protocol. This


503 robustness is especially relevant as multi-environment trials are common in variety testing (Isik


504 et al., 2017) and typically consist of trials spread over distant locations that do not have access to


505
the same equipment and field personnel. Therefore, image-based phenotyping enables the


506
quantification of flowering traits in an objective manner, improving the comparability of data


507 between trials conducted under various conditions.


508
The three-parameter logistic model used to describe the flowering phase, and the linear model


509
applied to the abscission phase showed good model fits, confirming their suitability for capturing


510
major developmental trends. In contrast, poorer model fits were more frequently observed for the


511
abscission phase for the ETH23 trial. This was associated with a second flowering peak in some


512
plots towards the end of August 2023, triggered by a specific weather pattern characterized by a


513
prolonged period of hot and dry weather followed by five days of heavy rain (Supplementary Figure


514 3). No comparable environmental conditions were observed in summer 2024 (Supplementary


515
Figure 4) and no second flowering peak occurred. This indicates a limitation of the modelling


516 approach rather than data quality _per se_, as the assumption of a single linear trajectory does not hold


517 true under all environmental conditions. Consequently, future studies based on larger datasets


518 collected across a wider range of environments may benefit from more flexible modelling approaches


519 to better capture complex flowering dynamics.


520
### 4.1. Image-based, high throughput phenotyping of buckwheat flowering 521 patterns


522 Vegetative development stages assessed using the BBCH scale were most strongly correlated


523
with Start Flowering DAS, the image-derived flowering trait that is temporally closest to vegetative


524 stage. This correlation was strongest for later vegetative phases (e.g., formation of the fourth true


525 leaf) and weaker for earlier stages such as cotyledon formation. This pattern is expected, as the


526
image-based analysis focused on floral development rather than vegetative growth. Thus, our


527
imaging approach complements efforts to optimise BBCH scales for buckwheat (Arduini et al.,


528
2016; Grimes et al., 2025) by providing an automated workflow for assessing flowering dynamics


529
in large scale field trials and variety testing. In particular, it delivers high-quality phenotypic data


[bioRxiv preprint doi: https://doi.org/10.64898/2026.11.03.763633; this version posted November 4, 2026. The copyright holder for this](https://doi.org/10.64898/2026.11.03.763633)

preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in
[perpetuity. It is made available under a CC-BY 4.0 International license.](http://creativecommons.org/licenses/by/4.0/)


530
for the period from the onset of flowering to the early seed set, which has been described as the


531 most critical period for seed formation (Guglielmini et al., 2019).


532
In addition to timing of flowering, the image-based trait detection provides quantitative


533
information on flowering intensity, which is not captured by the BBCH scale and, to our


534
knowledge, not previously assessed under field conditions. Traits related to flowering intensity


535 have been shown to be relevant for plant breeding in other crops (Zhang et al., 2020, Oppliger et


536 al., 2025), and combining Flowering AUC with the yield data could serve as a proxy for expected


537
yield. Most promising would be the development of a metric to assess seed set efficiency (i.e., the


538
proportion of flowers that produce seeds) under field conditions. Although seed set efficiency is


539 considered a major limitation for buckwheat productivity (Campbell, 1997), its assessment has


540
been constrained by the difficultly of quantifying flower numbers in the field. This limitation could


541
be addressed by adapting the methods described here. Beyond seed formation, flowering


542
intensity is also relevant if the plant is used in ecological flower strips (Strahm et al., 2019) or as a


543
melliferous crop (Halbrecq et al., 2005), further illustrating the broader relevance of flowering


544 intensity traits.

### 545 4.2. Variation of phenology in the buckwheat collection


546
Very little data on the flowering behaviour of different buckwheat accessions is available.


547 However, Campbell (1997) reported data from variety testing in India (number of accessions =


548 408), Nepal (𝑛 = 309) and China (𝑛 = 964), including observations on the start of flowering, full


549
flowering and maturity. Comparison between these data and our results must be made with


550
caution due to differences in the definitions of development stages and the lack of metadata


551 about the trials, such as environmental conditions at the trial sites and information on sowing


552
time. Strahm et al. (2019) reported the start and full flowering of 14 accessions tested in


553 Switzerland under similar climatic conditions to those of our trials. The mean values reported for


554
start of flowering, full flowering and maturity are generally comparable to those found here.


555
However, our data cover a wider range of values, which likely reflects the greater genetic diversity


556 of the accessions investigated.


557
The diversity of flowering behaviours contained in our buckwheat collection represents a resource


558
to be harnessed to breed new buckwheat varieties tailored to fulfil specific roles in the crop


559 rotation. In the literature, buckwheat has been suggested as a main or second crop for grain


560 production (Radics and Mikóházi, 2010), a fodder crop (Amelchanka et al., 2010), a green manure


561
(Taranenko et al., 2016) or a melliferous flower crop (Halbrecq et al., 2005). Each of these roles


562
requires distinct flowering characteristics. For example, varieties intended as a main crop for grain


[bioRxiv preprint doi: https://doi.org/10.64898/2026.11.03.763633; this version posted November 4, 2026. The copyright holder for this](https://doi.org/10.64898/2026.11.03.763633)

preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in
[perpetuity. It is made available under a CC-BY 4.0 International license.](http://creativecommons.org/licenses/by/4.0/)


563 production should be bred to maximize seed set, whereas varieties developed for green manure


564
may benefit from limited seed set, to reduce volunteer plants in subsequent crops (McKenzie

565 Gopsill, 2025). Varieties used as a second crop must exhibit rapid development and reach


566
maturity before the first frost, while melliferous varieties would be selected for sustained flower


567 production (i.e., maximize Flowering AUC or Flower Amplitude). Across all applications, a rapid


568 and homogeneous abscission phase (i.e., steep Abscission Rate and short Abscission Duration)


569 would make harvesting and processing easier (Funatsuki et al., 2000; Arduini et al., 2016).


570
However, our results also show that sowing time (i.e., different day length during development)


571
has a large effect on buckwheat flowering behaviour as previously reported (Hagiwara et al., 1998;


572 Michiyama et al., 2005; Arduini et al., 2016). Therefore, it is crucial to assess the response of


573
different accessions to changing day length and to perform variety testing and selection under


574 environmental conditions similar to their intended production location and system (Ewing et al.,


575 2019).

### 576 4.3. From pictures to genes


577
In total, 343 unique markers were associated with one or more of the 22 image-derived flowering


578
phenology traits, reflecting the complexity of such traits as reported for other crops (C. Liu et al.,


579 2021; Shi et al., 2022; Y. Li et al., 2023). Of these associations, 171 occurred within annotated


580
genes, offering insights into the possible molecular mechanisms underlying the variation


581
detected. The regional profiles generated for selected gene-overlapping markers showed


582 localized association peaks and no broad continuous high-LD blocks across the neighbouring


583 markers. This pattern is consistent with the rapid LD decay expected in outcrossing species and


584 with previous reports of low LD in allogamous common buckwheat populations (Flint-Garcia et


585 al., 2003; Yabe et al., 2018; Matsui and Yasui, 2020). LD was calculated from genome-wide allele

586 frequency fingerprints, meaning that the local 𝑟 [2] heatmaps describe correlations among


587
accession-level allele-frequency profiles rather than individual haplotypes (Nay et al., 2020).


588 These gene-overlapping associations, together with their functional annotations and regional


589
GWAS profiles, highlight candidates that could be considered for different breeding objectives.


590 For example, a marker associated with Flowering Start DAS lies within a DPH1 homolog and is


591 annotated as introducing a premature stop codon. _Arabidopsis dph1_ mutants show delayed


592 development and reduced cellular proliferation (H. Zhang et al., 2022). Similarly, the marker


593 associated with Mid Flowering mapped within an auxin response factor 2 (ARF2) homolog. In


594 _Arabidopsis_, _arf2_ mutants show altered flowering time (Ellis et al., 2005; Okushima et al., 2005),


595
supporting this locus as a plausible regulator of flowering progression. Together, these loci


[bioRxiv preprint doi: https://doi.org/10.64898/2026.11.03.763633; this version posted November 4, 2026. The copyright holder for this](https://doi.org/10.64898/2026.11.03.763633)

preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in
[perpetuity. It is made available under a CC-BY 4.0 International license.](http://creativecommons.org/licenses/by/4.0/)


596 represent promising candidates for marker-assisted selection or introgression aimed at adjusting


597
flowering onset or synchrony.


598 In contrast, the marker associated with Flowering Amplitude is directly relevant for breeding


599
melliferous ideotypes, as it directly affects the flowering intensity during the growing season. This


600
marker lies within a MDN1 homolog, which affects flowering time and auxin homeostasis in weak

601 _Arabidopsis_ mutants (K. Li et al., 2021, 2022), making it a candidate for flower development and


602 therefore nectar availability. Finally, the marker associated with Abscission Rate is relevant for


603
ideotypes targeting the persistence versus rapid decline of floral display after peak flowering. This


604 locus lies within a homolog of an ethylene signalling gene (EIN2) in _Arabidopsis_ . This gene is


605 associated with floral organ abscission and senescence phenotypes in _Arabidopsis_ and tobacco


606
(Patterson and Bleecker, 2004; Chakrabarti and Bharti, 2023), suggesting a role in flower retention


607 dynamics. Depending on the breeding target, this association could be explored to either prolong


608
flowering (e.g. melliferous) or to promote a faster abscission (e.g., synchronize maturity, improve


609
management timing). While these highlighted examples illustrate specific markers associated


610
with flowering time, flower intensity and abscission, many additional associations exist within


611 each of these categories (Table S7). These results provide a catalogue of candidate loci and genes


612 that can now be prioritized for functional validation and converted into molecular tools to tailor


613
common buckwheat varieties for target flowering dynamics and end-use ideotypes.

## 614 5. Conclusion


615
This work illustrates how image-based phenotyping coupled with statistical modelling can offer a


616
robust and scalable platform to detect flowering traits in buckwheat across different


617 environments, divergent varieties and phenotyping setups, including low-cost imaging devices


618 such as smartphones and digital cameras. The resulting traits have high heritability and


619
repeatability and correlate highly with manually measured BBCH flowering stages validating their


620
reliability. Moreover, the ability to quantify flowering intensity adds a complementary layer of


621 information not available in BBCH data. Thus, our phenotyping approach has the potential to


622
accelerate buckwheat breeding and improve its flowering behaviour, overcoming a major hurdle


623 to the crop’s adoption by farmers.


624
## Funding


625 C.Q.T. and F.H. were supported by the European Union’s Horizon 2020 research and innovation


626 program under the Marie Skłodowska-Curie grant agreement No 847585 – RESPONSE and by the


[bioRxiv preprint doi: https://doi.org/10.64898/2026.11.03.763633; this version posted November 4, 2026. The copyright holder for this](https://doi.org/10.64898/2026.11.03.763633)

preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in
[perpetuity. It is made available under a CC-BY 4.0 International license.](http://creativecommons.org/licenses/by/4.0/)


627
Federal Office for Agriculture (FOAG) within the framework of the National Plan of Action for the


628 Conservation and Sustainable Use of Plant Genetic Resources for Food and Agriculture (NPA

629 PGRFA).


630
## Conflicts of interest


631 The authors declare that the research was conducted in the absence of any commercial or


632
financial relationships that could be construed as a potential conflict of interest.


633
## Data availability


634 Code and data will be made available upon reasonable request.

## 635 Authors contributions


636
F.H., L.K., S.Y. conceived this study and B.S. acquired funding. F.H. designed the field trials and


637 managed the trials at ETH, took images (ETH only), contributed to statistical analysis and wrote


638 the manuscript. C.Q.T, designed and implemented the image analysis pipeline, compiled the


639 training dataset, contributed to statistical analysis, performed GWAS and wrote the manuscript.


640
F.T. provided the scripts for the modelling of flowering dynamics. G.R. managed the trials at LfL


641 and took images. S.Y., F.T., L.K. and B.S. edited the manuscript.


642
## Acknowledgements


643
We would like to sincerely thank Heidi Heuberger for providing administrative and scientific


644 support to the project as the responsible coordinator of the trial at LfL. Additionally, we would like


645
to thank Simon Corrado for his support with the establishment and management of the field trials


646 at ETH as well as our interns and student assistants Shuyo Nodasaka, Benjamim Fabri and Manuel


647
Cavigelli for their work in maintaining field trials and taking images.

## 648 References


649 Amelchanka, S. L., Kreuzer, M., & Leiber, F. (2010). Utility of buckwheat ( _Fagopyrum esculentum_


650 _in vitro_ ruminal fermentation and
Moench) as feed: Effects of forage and grain on


[bioRxiv preprint doi: https://doi.org/10.64898/2026.11.03.763633; this version posted November 4, 2026. The copyright holder for this](https://doi.org/10.64898/2026.11.03.763633)

preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in
[perpetuity. It is made available under a CC-BY 4.0 International license.](http://creativecommons.org/licenses/by/4.0/)


651 performance of dairy cows. _Animal Feed Science and Technology_, _155_ (2), 111–121.


652 https://doi.org/10.1016/j.anifeedsci.2009.10.007


653 Arduini, I., Masoni, A., & Mariotti, M. (2016). A growth scale for the phasic development of


654 common buckwheat. _Acta Agriculturae Scandinavica, Section B — Soil & Plant Science_,


655 _66_ (3), 215–228. https://doi.org/10.1080/09064710.2015.1087587


656 Campbell, C. G. (1997). _Buckwheat,_ _Fagopyrum_ _esculentus_ _Moench_ .


657 https://hdl.handle.net/10568/104279


658 Cantalapiedra, C. P., Hernández-Plaza, A., Letunic, I., Bork, P., & Huerta-Cepas, J. (2021).


659 eggNOG-mapper v2: Functional Annotation, Orthology Assignments, and Domain


660 Prediction at the Metagenomic Scale. _Molecular Biology and Evolution_, _38_ (12), 5825–


661 5829. https://doi.org/10.1093/molbev/msab293


662 Cawoy, V., Ledent, J.-F., Kinet, J.-M., & Jacquemart, A.-L. (2009). Floral biology of common


663 buckwheat (Fagopyrum esculentum Moench). _The European Journal of Plant Science and_


664 _Biotechnology_, _3_ (1), 1–9.


665 Chakrabarti, M., & Bharti, S. (2023). Role of EIN2-mediated ethylene signaling in regulating petal


666 senescence, abscission, reproductive development, and hormonal crosstalk in tobacco.


667 _Plant Science_, _332_, 111699. https://doi.org/10.1016/j.plantsci.2023.111699


668 Cobb, J. N., DeClerck, G., Greenberg, A., Clark, R., & McCouch, S. (2013). Next-generation


669 phenotyping: Requirements and strategies for enhancing our understanding of genotype–


670 phenotype relationships and its relevance to crop improvement. _Theoretical and Applied_


671 _Genetics_, _126_ (4), 867–887. https://doi.org/10.1007/s00122-013-2066-0


672 Cullis, B. R., Smith, A. B., & Coombes, N. E. (2006). On the design of early generation variety trials


673 with correlated data. _Journal of Agricultural, Biological, and Environmental Statistics_,


674 _11_ (4), 381–393. https://doi.org/10.1198/108571106X154443


675 Dutta, A., & Zisserman, A. (2019). The VIA Annotation Software for Images, Audio and Video.


676 _Proceedings of the 27th ACM International Conference on Multimedia, MM ’19_, 2276–


677 2279. https://doi.org/10.1145/3343031.3350535


[bioRxiv preprint doi: https://doi.org/10.64898/2026.11.03.763633; this version posted November 4, 2026. The copyright holder for this](https://doi.org/10.64898/2026.11.03.763633)

preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in
[perpetuity. It is made available under a CC-BY 4.0 International license.](http://creativecommons.org/licenses/by/4.0/)


678 Ellis, C. M., Nagpal, P., Young, J. C., Hagen, G., Guilfoyle, T. J., & Reed, J. W. (2005). AUXIN


679
RESPONSE FACTOR1 and AUXIN RESPONSE FACTOR2regulate senescence and floral


680 organ abscission in Arabidopsis thaliana. _Development_, _132_ (20), 4563–4574.


681 https://doi.org/10.1242/dev.02012


682
Ewing, P. M., Runck, B. C., Kono, T. Y. J., & Kantar, M. B. (2019). The home field advantage of modern


683 plant breeding. _PLOS_ _ONE_, _14_ (12), e0227079.


684 https://doi.org/10.1371/journal.pone.0227079


685 Farooq, S., Rehman, R. U., Pirzadah, T. B., Malik, B., Dar, F. A., & Tahir, I. (2016). Cultivation,


686 Agronomic Practices, and Growth Performance of Buckwheat. In M. Zhou, I. Kreft, S.-H.


687 Woo, N. Chrungoo, & G. Wieslander (Eds), _Molecular Breeding and Nutritional Aspects of_


688 _Buckwheat_ (pp. 299–319). Academic Press. https://doi.org/10.1016/B978-0-12-803692

689 1.00023-7


690 Fisher, R. A. (1992). Statistical Methods for Research Workers. In S. Kotz & N. L. Johnson (Eds),


691 _Breakthroughs in Statistics: Methodology and Distribution_ (pp. 66–70). Springer.


692 https://doi.org/10.1007/978-1-4612-4380-9_6


693 Flint-Garcia, S. A., Thornsberry, J. M., & Buckler, E. S. (2003). Structure of Linkage Disequilibrium


694 in Plants. _Annual_ _Review_ _of_ _Plant_ _Biology_, _54_ (1), 357–374.


695 https://doi.org/10.1146/annurev.arplant.54.031902.134907


696 Funatsuki, H., Maruyama-Funatsuki, W., Fujino, K., & Agatsuma, M. (2000). Ripening Habit of


697 Buckwheat. _Crop_ _Science_, _40_ (4), 1103–1108.


698 https://doi.org/10.2135/cropsci2000.4041103x


699 Giménez-Gallego, J., Martinez-del-Rincon, J., González-Teruel, J. D., Navarro-Hellín, H., Navarro,


700 P. J., & Torres-Sánchez, R. (2024). On-tree fruit image segmentation comparing Mask R

701 CNN and Vision Transformer models. Application in a novel algorithm for pixel-based fruit


702 size estimation. _Computers_ _and_ _Electronics_ _in_ _Agriculture_, _222_, 109077.


703 https://doi.org/10.1016/j.compag.2024.109077


[bioRxiv preprint doi: https://doi.org/10.64898/2026.11.03.763633; this version posted November 4, 2026. The copyright holder for this](https://doi.org/10.64898/2026.11.03.763633)

preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in
[perpetuity. It is made available under a CC-BY 4.0 International license.](http://creativecommons.org/licenses/by/4.0/)


704
Grimes, S., Baier, C., Hartung, J., Longin, C., & Graeff-Hönninger, S. (2025). Phenological growth


705 stages of common buckwheat (Fagopyrum esculentum Moench): An extended BBCH


706 scale. _Journal Für Kulturpflanzen_, _77_ (01), 89–100. https://doi.org/10.5073/JfK.2025.01.11


707 Guglielmini, A. C., Forcat, J. I., & Miralles, D. J. (2019). The critical period for yield determination in


708 common buckwheat ( _Fagopyrum esculentum_ Moench). _European Journal of Agronomy_,


709 _110_, 125933. https://doi.org/10.1016/j.eja.2019.125933


710
Hagiwara, M., Inoue, N., & Matano, T. (1998). Variability in the Length of Flower Bud Differentiation


711 Period of Common Buckwheat—Papers—Researchmap. _Fagopyrum_, _15_, 55–64.


712
Halbrecq, B., Romedenne, P., & Ledent, J. F. (2005). Evolution of flowering, ripening and seed set


713 in buckwheat ( _Fagopyrum esculentum_ Moench): Quantitative analysis. _European Journal_


714 _of Agronomy_, _23_ (3), 209–224. https://doi.org/10.1016/j.eja.2004.11.006


715 He, K., Gkioxari, G., Dollár, P., & Girshick, R. (2017). Mask R-CNN. _2017 IEEE International_


716 _Conference_ _on_ _Computer_ _Vision_ _(ICCV)_, 2980–2988.


717 https://doi.org/10.1109/ICCV.2017.322


718
Hess, F., Chen, Y., López, M.-E., Colliquet, A., Stoffel-Studer, I., Mac, V., Grob, S., Kölliker, R., &


719 Studer, B. (2026a). Near-complete, haplotype-resolved genome assembly of common


720 buckwheat (Fagopyrum esculentum Moench). _Scientific_ _Data_ .


721 https://doi.org/10.1038/s41597-026-08092-4


722 Hess, F., Kronenberg, L., Rimat, G., Heuberger, H., Herrmann, M. H., Benyr, V., Yates, S., & Studer,


723 B. (2026b). _Agronomic characterisation of a common buckwheat germplasm collection_


724 _unlocks_ _resources_ _for_ _genomics-assisted_ _ideotype_ _breeding_ . Plant Biology.


725 https://doi.org/10.64898/2026.10.09.757874


726 Hess, F., Kronenberg, L., Yates, S., Ariza-Suarez, D., Nay, M. M., Heuberger, H., Büttner, B., Zeller,


727 F. J., Eigenmann, C., Walter, A., & Studer, B. (2026c). Genetic characterization of a global


728 collection of common buckwheat to harness genetic resources for agricultural diversity.


729 _Frontiers in Plant Science_, _17_ . https://doi.org/10.3389/fpls.2026.1902211


[bioRxiv preprint doi: https://doi.org/10.64898/2026.11.03.763633; this version posted November 4, 2026. The copyright holder for this](https://doi.org/10.64898/2026.11.03.763633)

preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in
[perpetuity. It is made available under a CC-BY 4.0 International license.](http://creativecommons.org/licenses/by/4.0/)


730 _Theoretical and_
Hill, W. G., & Robertson, A. (1968). Linkage disequilibrium in finite populations.


731 _Applied Genetics_, _38_ (6), 226–231. https://doi.org/10.1007/BF01245622


732 Huang, M., Liu, X., Zhou, Y., Summers, R. M., & Zhang, Z. (2019). BLINK: A package for the next


733 level of genome-wide association studies with both individuals and markers in the


734 millions. _GigaScience_, _8_ (2), giy154. https://doi.org/10.1093/gigascience/giy154


735 Huerta-Cepas, J., Szklarczyk, D., Heller, D., Hernández-Plaza, A., Forslund, S. K., Cook, H., Mende,


736 D. R., Letunic, I., Rattei, T., Jensen, L. J., von Mering, C., & Bork, P. (2019). eggNOG 5.0: A


737 hierarchical, functionally and phylogenetically annotated orthology resource based on


738 5090 organisms and 2502 viruses. _Nucleic Acids Research_, _47_ (D1), D309–D314.


739 https://doi.org/10.1093/nar/gky1085


740 Isik, F., Holland, J., & Maltecca, C. (2017). Multi Environmental Trials. In F. Isik, J. Holland, & C.


741 Maltecca (Eds), _Genetic Data Analysis for Plant and Animal Breeding_ (pp. 227–262).


742 Springer International Publishing. https://doi.org/10.1007/978-3-319-55177-7_8


743 Joshi, D. C., Chaudhari, G. V., Sood, S., Kant, L., Pattanayak, A., Zhang, K., Fan, Y., Janovská, D.,


744 Meglič, V., & Zhou, M. (2019). Revisiting the versatile buckwheat: Reinvigorating genetic


745 gains through integrated breeding and genomics approach. _Planta_, _250_ (3), 783–801.


746 https://doi.org/10.1007/s00425-018-03080-4


747
Lex, A., Gehlenborg, N., Strobelt, H., Vuillemot, R., & Pfister, H. (2014). UpSet: Visualization of


748 Intersecting Sets. _IEEE Transactions on Visualization and Computer Graphics_, _20_ (12),


749 1983–1992. https://doi.org/10.1109/TVCG.2014.2346248


750 Li, K., Wang, P., Ding, T., Hou, L., Li, G., Zhao, C., Zhao, S., Wang, X., Li, P., Li, K., Wang, P., Ding, T.,


751 Hou, L., Li, G., Zhao, C., Zhao, S., Wang, X., & Li, P. (2022). Mutation of an Essential 60S


752 Ribosome Assembly Factor MIDASIN 1 Induces Early Flowering in Arabidopsis.


753 _International_ _Journal_ _of_ _Molecular_ _Sciences_, _23_ (12).


754 https://doi.org/10.3390/ijms23126509


755 Li, K., Zhou, X., Sun, X., Li, G., Hou, L., Zhao, S., Zhao, C., Ma, C., Li, P., & Wang, X. (2021).


756 Coordination between MIDASIN 1-mediated ribosome biogenesis and auxin modulates


[bioRxiv preprint doi: https://doi.org/10.64898/2026.11.03.763633; this version posted November 4, 2026. The copyright holder for this](https://doi.org/10.64898/2026.11.03.763633)

preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in
[perpetuity. It is made available under a CC-BY 4.0 International license.](http://creativecommons.org/licenses/by/4.0/)


757 plant development. _Journal_ _of_ _Experimental_ _Botany_, _72_ (7), 2501–2513.


758 https://doi.org/10.1093/jxb/erab025


759 Li, Y., Guo, L., Wang, Z., Zhao, D., Guo, D., Carlson, J. E., Yin, W., & Hou, X. (2023). Genome-wide


760
association study of 23 flowering phenology traits and 4 floral agronomic traits in tree


761
peony (Paeonia section Moutan DC.) reveals five genes known to regulate flowering time.


762 _Horticulture Research_, _10_ (2), uhac263. https://doi.org/10.1093/hr/uhac263


763 Lin, T.-Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollár, P., & Zitnick, C. L. (2014).


764 Microsoft COCO: Common Objects in Context. In D. Fleet, T. Pajdla, B. Schiele, & T.


765 Tuytelaars (Eds), _Computer Vision – ECCV 2014_ (pp. 740–755). Springer International


766 Publishing. https://doi.org/10.1007/978-3-319-10602-1_48


767 Liu, C., Tu, Y., Liao, S., Fu, X., Lian, X., He, Y., Xie, W., & Wang, G. (2021). Genome-wide association


768
study of flowering time reveals complex genetic heterogeneity and epistatic interactions


769 in rice. _Gene_, _770_, 145353. https://doi.org/10.1016/j.gene.2020.145353


770 Liu, X., Huang, M., Fan, B., Buckler, E. S., & Zhang, Z. (2016). Iterative Usage of Fixed and Random


771 _PLOS_
Effect Models for Powerful and Efficient Genome-Wide Association Studies.


772 _Genetics_, _12_ (2), e1005767. https://doi.org/10.1371/journal.pgen.1005767


773 Matsui, K., & Yasui, Y. (2020). Buckwheat heteromorphic self-incompatibility: Genetics, genomics


774 and application to breeding. _Breeding_ _Science_, _70_ (1), 32–38.


775 https://doi.org/10.1270/jsbbs.19083


776 McKenzie-Gopsill, A. (2025). The critical period of cover crop management: A framework for


777 maximizing biomass potential and minimizing volunteers with buckwheat. _Agronomy_


778 _Journal_, _117_ (3), e70097. https://doi.org/10.1002/agj2.70097


779
Michiyama, H., Tsuchimoto, K., Tani, K., Hirano, T., Hayashi, H., & Campbell, C. (2005). Influence


780 of Day Length on Stem Growth, Flowering, Morphology of Flower Clusters, and Seed-Set


781 in Buckwheat (Fagopyrum esculentum Moench). _Plant Production Science_, _8_ (1), 44–50.


782 https://doi.org/10.1626/pps.8.44


[bioRxiv preprint doi: https://doi.org/10.64898/2026.11.03.763633; this version posted November 4, 2026. The copyright holder for this](https://doi.org/10.64898/2026.11.03.763633)

preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in
[perpetuity. It is made available under a CC-BY 4.0 International license.](http://creativecommons.org/licenses/by/4.0/)


783 Morishita, T., Hara, T., & Hara, T. (2020). Important agronomic characteristics of yielding ability in


784
common buckwheat; ecotype and ecological differentiation, preharvest sprouting


785 resistance, shattering resistance, and lodging resistance. _Breeding Science_, _70_ (1), 39–47.


786 https://doi.org/10.1270/jsbbs.19020


787 Murphy, K. M., Ludwig, E., Gutierrez, J., & Gehan, M. A. (2024). Deep Learning in Image-Based Plant


788 Phenotyping. _Annual Review of Plant Biology_, _75_ (Volume 75, 2024), 771–795.


789 https://doi.org/10.1146/annurev-arplant-070523-042828


790
Nature Communications Editorial Staff. (2024). Efficient genetic improvement of orphan crops


791 cannot follow the old path. _Nature_ _Communications_, _15_ (1), 321.


792 https://doi.org/10.1038/s41467-023-44458-7


793 Nay, M. M., Byrne, S. L., Pérez, E. A., Walter, A., & Studer, B. (2020). Genetic characterization of


794
buckwheat accessions through genome-wide allele frequency fingerprints / Genetska


795 karakterizacija vzorcev ajde z odtisi frekvence alelov v genomu. _Folia Biologica et_


796 _Geologica_, _61_ (1), 17–23. https://doi.org/10.3986/fbg0063


797 Oakey, H., Verbyla, A., Pitchford, W., Cullis, B., & Kuchel, H. (2006). Joint modeling of additive and


798 non-additive genetic line effects in single field trials. _Theoretical and Applied Genetics_,


799 _113_ (5), 809–819. https://doi.org/10.1007/s00122-006-0333-z


800 Ohsawa, R. (2020). Current status and prospects of common buckwheat breeding in Japan.


801 _Breeding Science_, _70_ (1), 3–12. https://doi.org/10.1270/jsbbs.19108


802 Okushima, Y., Mitina, I., Quach, H. L., & Theologis, A. (2005). AUXIN RESPONSE FACTOR 2 (ARF2):


803 A pleiotropic developmental regulator. _The_ _Plant_ _Journal_, _43_ (1), 29–46.


804 https://doi.org/10.1111/j.1365-313X.2005.02426.x


805 Oppliger, C., Zenkl, R., Walter, A., & Keller, B. (2025). Investigating pea ( _Pisum sativum_ L.) flowering


806 with high throughput field phenotyping and object detection. _Smart Agricultural_


807 _Technology_, _11_, 100942. https://doi.org/10.1016/j.atech.2025.100942


[bioRxiv preprint doi: https://doi.org/10.64898/2026.11.03.763633; this version posted November 4, 2026. The copyright holder for this](https://doi.org/10.64898/2026.11.03.763633)

preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in
[perpetuity. It is made available under a CC-BY 4.0 International license.](http://creativecommons.org/licenses/by/4.0/)


808 Paril, J., Cogan, N. O. I., & Malmberg, M. M. (2025). Imputef: Imputation of polyploid genotype


809 classes and allele frequencies. _BMC_ _Genomics_, _26_ (1), 946.


810 https://doi.org/10.1186/s12864-025-12141-4


811 Patterson, S. E., & Bleecker, A. B. (2004). Ethylene-Dependent and -Independent Processes


812 Associated with Floral Organ Abscission in Arabidopsis. _Plant Physiology_, _134_ (1), 194–


813 203. https://doi.org/10.1104/pp.103.028027


814 Radics, L., & Mikóházi, D. (2010). Principles of Common Buckwheat Production. _The European_


815 _Journal of Plant Science and Biotechnology_, (1), 57–63.


816 Ren, S., He, K., Girshick, R., & Sun, J. (2017). Faster R-CNN: Towards Real-Time Object Detection


817 with Region Proposal Networks. _IEEE Transactions on Pattern Analysis and Machine_


818 _Intelligence_, _39_ (6), 1137–1149. https://doi.org/10.1109/TPAMI.2016.2577031


819 Rodríguez-Álvarez, M. X., Boer, M. P., van Eeuwijk, F. A., & Eilers, P. H. C. (2018). Correcting for


820 spatial heterogeneity in plant breeding experiments with P-splines. _Spatial Statistics_, _23_,


821 52–71. https://doi.org/10.1016/j.spasta.2017.10.003


822 Roth, L., Kronenberg, L., Aasen, H., Walter, A., Hartung, J., van Eeuwijk, F., Piepho, H.-P., & Hund,


823
A. (2024). High-throughput field phenotyping reveals that selection in breeding has


824
affected the phenology and temperature response of wheat in the stem elongation phase.


825 _Journal of Experimental Botany_, _75_ (7), 2084–2099. https://doi.org/10.1093/jxb/erad481


826 Roth, L., Rodríguez-Álvarez, M. X., van Eeuwijk, F., Piepho, H.-P., & Hund, A. (2021). Phenomics


827 data processing: A plot-level model for repeated measurements to extract the timing of


828 key stages and quantities at defined time points. _Field Crops Research_, _274_, 108314.


829 https://doi.org/10.1016/j.fcr.2021.108314


830 Segura, V., Vilhjálmsson, B. J., Platt, A., Korte, A., Seren, Ü., Long, Q., & Nordborg, M. (2012). An


831
efficient multi-locus mixed-model approach for genome-wide association studies in


832 structured populations. _Nature_ _Genetics_, _44_ (7), 825–830.


833 https://doi.org/10.1038/ng.2314


[bioRxiv preprint doi: https://doi.org/10.64898/2026.11.03.763633; this version posted November 4, 2026. The copyright holder for this](https://doi.org/10.64898/2026.11.03.763633)

preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in
[perpetuity. It is made available under a CC-BY 4.0 International license.](http://creativecommons.org/licenses/by/4.0/)


834 Shi, J., Wang, Y., Wang, C., Wang, L., Zeng, W., Han, G., Qiu, C., Wang, T., Tao, Z., Wang, K., Huang,


835 S., Yu, S., Wang, W., Chen, H., Chen, C., He, C., Wang, H., Zhu, P., Hu, Y., … Li, P. (2022).


836 Linkage mapping combined with GWAS revealed the genetic structural relationship and


837 candidate genes of maize flowering time-related traits. _BMC Plant Biology_, _22_ (1), 328.


838 https://doi.org/10.1186/s12870-022-03711-9


839 Škrabanja, V., & Kreft, I. (2016). Nutritional Value of Buckwheat Proteins and Starch. In _Molecular_


840 _Breeding and Nutritional Aspects of Buckwheat_ (pp. 169–176). Academic Press.


841 https://doi.org/10.1016/B978-0-12-803692-1.00013-4


842 Strahm, S., Füglistaller, D., Lädrach, C., Enggist, A., Thuet, A., Luginbühl, C., Ramseier, H., &


843 Hiltbrunner, J. (2019). Buchweizenanbau in der Schweiz: Neue Sorten für eine alte


844 Nischenkultur. _Agrarforschung_ _Schweiz_, _10_ (5), 198–205.


845 https://doi.org/10.24451/arbor.8334


846 Taranenko, L. K., Yatsyshen, O. L., Taranenko, P. P., & Taranenko, T. P. (2016). The Unique Value of


847 Buckwheat as a Most Important Traditional Cereal Crop in Ukraine. In _Molecular Breeding_


848 _and_ _Nutritional_ _Aspects_ _of_ _Buckwheat_ (pp. 81–85). Academic Press.


849 https://doi.org/10.1016/B978-0-12-803692-1.00006-7


850 Tschurr, F., Roth, L., Storni, N., Zumsteg, O., Walter, A., & Anderegg, J. (2024). Temporal resolution


851 trumps spectral resolution in UAV-based monitoring of cereal senescence dynamics.


852 _Plant Methods_, _20_ (1), 188. https://doi.org/10.1186/s13007-024-01308-x


853 VanRaden, P. M. (2007). Genomic measures of relationship and inbreeding. _Interbull Bulletin_, (37),


854 33–33.


855 Varshney, R. K., Ribaut, J.-M., Buckler, E. S., Tuberosa, R., Rafalski, J. A., & Langridge, P. (2012).


856 Can genomics boost productivity of orphan crops? _Nature Biotechnology_, _30_ (12), 1172–


857 1176. https://doi.org/10.1038/nbt.2440


858 Wang, J., & Zhang, Z. (2021). GAPIT Version 3: Boosting Power and Accuracy for Genomic


859 Association and Prediction. _Genomics, Proteomics & Bioinformatics_, _19_ (4), 629–640.


860 https://doi.org/10.1016/j.gpb.2021.08.005


[bioRxiv preprint doi: https://doi.org/10.64898/2026.11.03.763633; this version posted November 4, 2026. The copyright holder for this](https://doi.org/10.64898/2026.11.03.763633)

preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in
[perpetuity. It is made available under a CC-BY 4.0 International license.](http://creativecommons.org/licenses/by/4.0/)


861 Wang, X., Xuan, H., Evers, B., Shrestha, S., Pless, R., & Poland, J. (2019). High-throughput


862
phenotyping with deep learning gives insight into the genetic architecture of flowering time


863 in wheat. _GigaScience_, _8_ (11), giz120. https://doi.org/10.1093/gigascience/giz120


864 Yabe, S., Hara, T., Ueno, M., Enoki, H., Kimura, T., Nishimura, S., Yasui, Y., Ohsawa, R., & Iwata, H.


865 (2018). Potential of Genomic Selection in Mass Selection Breeding of an Allogamous


866 Crop: An Empirical Study to Increase Yield of Common Buckwheat. _Frontiers in Plant_


867 _Science_, _9_, 276. https://doi.org/10.3389/fpls.2018.00276


868 Yabe, S., & Iwata, H. (2020). Genomics-assisted breeding in minor and pseudo-cereals. _Breeding_


869 _Science_, _70_ (1), 19–31. https://doi.org/10.1270/jsbbs.19100


870 Yu, J., Pressoir, G., Briggs, W. H., Vroh Bi, I., Yamasaki, M., Doebley, J. F., McMullen, M. D., Gaut, B.


871
S., Nielsen, D. M., Holland, J. B., Kresovich, S., & Buckler, E. S. (2006). A unified mixed

872 model method for association mapping that accounts for multiple levels of relatedness.


873 _Nature Genetics_, _38_ (2), 203–208. https://doi.org/10.1038/ng1702


874 Zhang, C., Craine, W. A., McGee, R. J., Vandemark, G. J., Davis, J. B., Brown, J., Hulbert, S. H.,


875 Sankaran, S., Zhang, C., Craine, W. A., McGee, R. J., Vandemark, G. J., Davis, J. B., Brown,


876 J., Hulbert, S. H., & Sankaran, S. (2020). Image-Based Phenotyping of Flowering Intensity


877 in Cool-Season Crops. _Sensors_, _20_ (5). https://doi.org/10.3390/s20051450


878 Zhang, H., Quintana, J., Ütkür, K., Adrian, L., Hawer, H., Mayer, K., Gong, X., Castanedo, L.,


879
Schulten, A., Janina, N., Peters, M., Wirtz, M., Brinkmann, U., Schaffrath, R., & Krämer, U.


880
(2022). Translational fidelity and growth of Arabidopsis require stress-sensitive


881 diphthamide biosynthesis. _Nature_ _Communications_, _13_ (1), 4009.


882 https://doi.org/10.1038/s41467-022-31712-7


883
Zuffa, F., Jung, M., Yates, S., Quesada-Traver, C., Patocchi, A., Studer, B., & Dow, G. (2025).


884 Interannual Variation of Stomatal Traits Impacts the Environmental Responses of Apple


885 Trees. _Plant, Cell & Environment_, _48_ (3), 2478–2491. https://doi.org/10.1111/pce.15302


886


[bioRxiv preprint doi: https://doi.org/10.64898/2026.11.03.763633; this version posted November 4, 2026. The copyright holder for this](https://doi.org/10.64898/2026.11.03.763633)

preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in
[perpetuity. It is made available under a CC-BY 4.0 International license.](http://creativecommons.org/licenses/by/4.0/)


887
## Figures


888


889 **Figure 1.** Overview of the image-based flower detection phenotyping method and downstream analysis. A. depicts


890 details on data acquisition such as imaging device (i.e. Literal®, smartphone and camera), number of accessions,


891 timepoints and images taken in each of the trials. B. illustrates the image preprocessing pipeline which includes border


892 cropping and image splitting in quadrants. C. shows the image segmentation modelling strategy. For training, the


893 number of quadrants in the training (𝑛 𝑡𝑟𝑎𝑖𝑛 ) and validation (𝑛 𝑣𝑎𝑙 ) splits are shown. For prediction, the strategy of


894 predicting in the cropped image as well as in the quadrants with the subsequent mask overlaying and merging is


[bioRxiv preprint doi: https://doi.org/10.64898/2026.11.03.763633; this version posted November 4, 2026. The copyright holder for this](https://doi.org/10.64898/2026.11.03.763633)

preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in
[perpetuity. It is made available under a CC-BY 4.0 International license.](http://creativecommons.org/licenses/by/4.0/)


895 illustrated for one instance. In D. the two-phase flowering dynamics modelling is shown, where the three-parameter


896 logistic function describing flowering and the linear function describing abscission are shown in green and magenta,


897 respectively. The 𝑥 axis represents the timepoints and the 𝑦 axis the percentage of pixels in the image assigned as flower


898 (Flower %). In E. the steps of downstream analysis are illustrated.


899


900 **Figure 2.** Representation of the two-phase modelling strategy showing the three-parameter logistic function (green)


901 modelling the flowering phase and the linear model (magenta) modelling the abscission phase. Additionally, all the


902 parameters extracted and calculated from the fits are displayed in boxes.


[bioRxiv preprint doi: https://doi.org/10.64898/2026.11.03.763633; this version posted November 4, 2026. The copyright holder for this](https://doi.org/10.64898/2026.11.03.763633)

preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in
[perpetuity. It is made available under a CC-BY 4.0 International license.](http://creativecommons.org/licenses/by/4.0/)


903


904 **Figure 3.** Training and validation metrics for the buckwheat flower image segmentation model. (A) Evolution of box


905 (orange) and mask (green) losses over training for the training set (lines) and the validation checkpoints (points). (B)


906 Correlations between predicted and ground truth area across training checkpoints for the train (purple) and validation


907 (pink) sets for the raw mask predictions (circles) and the merged overlapping masks (triangles), note that the y-axis is


908 truncated for a better visualization. Green boxes in A and B depict the checkpoint at which the best metrics were


909 attained. (C) Scatterplot of Predicted Merged Area (%) and Ground Truth Area (%) for the selected best model depicting


910 both training (purple triangles) and validation (pink triangles) sets. (D) Box plot of predicted area (%) for images with a


911 ground truth area (%) equal to zero.


[bioRxiv preprint doi: https://doi.org/10.64898/2026.11.03.763633; this version posted November 4, 2026. The copyright holder for this](https://doi.org/10.64898/2026.11.03.763633)

preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in
[perpetuity. It is made available under a CC-BY 4.0 International license.](http://creativecommons.org/licenses/by/4.0/)


912


913 **Figure 4.** Flowering dynamics modelling. (A) Violin plots for the goodness-of-fit of both flowering phases at the field plot


914 level in the basis of 𝑅 [2] for the five used datasets: ETH23, ETH24, LfL23, LfL24 and ETH24L. Diamonds represent the

915 median of the distribution. Green and magenta represent the flowering and abscission phases, respectively. (B)


916 Example of model fits for the accession FE268 across all datasets. Different colours represent different plots for the

917 same accession. The solid lines represent the different Start Flowering DAS, the dashed lines the End Flowering DAS


918 and the dotted lines the Projected End of Abscission DAS for each plot


919


[bioRxiv preprint doi: https://doi.org/10.64898/2026.11.03.763633; this version posted November 4, 2026. The copyright holder for this](https://doi.org/10.64898/2026.11.03.763633)

preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in
[perpetuity. It is made available under a CC-BY 4.0 International license.](http://creativecommons.org/licenses/by/4.0/)


920


921 **Figure 5.** Effects of sowing time, accessions and environments on the flowering dynamics of buckwheat. (A1-3)

922 Boxplots showing the range of flowering traits within the ETH Zurich/BLW buckwheat collection sown as main (green;


923 𝑛= 167) or second crop (orange; 𝑛 = 51). The three plots show (1) traits related to time points in flower development,


924 (2) flowering intensity and (3) Flowering AUC. Stars indicate significant (𝑝 < 0.01, Wilcoxon-test) differences between


925 parameters compared. (B) Principal component analysis revealing differentiation of subclusters of buckwheat in terms

926 of flowering traits. For readability, only a selection of traits is shown in the figure, while the principal components are

927 based on all 22 image-derived flowering traits. Colours represent subclusters, and shapes growth habit. (C) Pairwise


928 comparisons in terms of Pearson’s correlation coefficient (𝑟) between environments of the 22 image-derived flowering

929 traits. The different line colours indicate traits related to duration (orange), intensity (blue) and timing (green) of


930 flowering phases, and the pink diamonds indicate the mean over all 22 traits per year-site comparison.


**Figure 6.** Pairwise Pearson correlations (𝑟) between BLUPs across years of ETH main experiments (A) within image-derived traits (B) and between image-derived and measured traits.


Positive correlations are depicted in green and negative correlations in orange, with a saturation and a square size proportional to their absolute correlation (|𝑟|). Asterisks indicate the


significance from two-sided correlation tests at different significance levels: [∗] 𝑝 < 0.05, [∗∗] 𝑝 < 0.01 and [∗∗∗] 𝑝 < 0.001. In (A) 𝑟̅ indicates the mean pairwise correlations within each group


excluding self-correlations .


[bioRxiv preprint doi: https://doi.org/10.64898/2026.11.03.763633; this version posted November 4, 2026. The copyright holder for this](https://doi.org/10.64898/2026.11.03.763633)

preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in
[perpetuity. It is made available under a CC-BY 4.0 International license.](http://creativecommons.org/licenses/by/4.0/)


**Figure 7.** Concordance of marker–trait associations between image-derived and measured flowering traits. (A) UpSet

plot showing unique and shared significant markers among four trait groups: Manual Abscission, Manual Flowering,


Image-derived Abscission and Image-derived Flower Period. The number of traits included in each group is indicated


by 𝑛 𝑡, the horizontal bars show the number of unique markers in each group, and the numbers at the left indicate the


total number of unique markers identified for each phenotyping method. Intersections involving both phenotyping

methods are classified according to whether the shared association concerns flowering traits (pink), abscission traits


(purple) or both trait categories (black). (B) Genomic distribution of significant marker–trait associations and


concordant signals across the eight chromosomes. Manually measured trait associations are shown to the left of each


chromosome and image-derived trait associations to the right. Point colour denotes the trait group, while point size is


proportional to Fisher’s combined −𝑙𝑜𝑔 10 (𝑝) for associations located at the same genomic position. Direct


associations denoted by horizontal bars connect identical SNP positions detected by both phenotyping methods. LD

supported associations denoted by squares indicate concordance between distinct SNP positions located within the


same LD block. Pink, purple and black denote associations involving only flowering traits, only abscission traits and


both trait categories, respectively.


[bioRxiv preprint doi: https://doi.org/10.64898/2026.11.03.763633; this version posted November 4, 2026. The copyright holder for this](https://doi.org/10.64898/2026.11.03.763633)

preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in
[perpetuity. It is made available under a CC-BY 4.0 International license.](http://creativecommons.org/licenses/by/4.0/)


[bioRxiv preprint doi: https://doi.org/10.64898/2026.11.03.763633; this version posted November 4, 2026. The copyright holder for this](https://doi.org/10.64898/2026.11.03.763633)

preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in
[perpetuity. It is made available under a CC-BY 4.0 International license.](http://creativecommons.org/licenses/by/4.0/)


**Figure 8.** Regional association patterns, linkage disequilibrium and allelic effects for four selected marker–trait


associations. (A) chr1_163426249_G/A associated with Mid Flowering. (B) chr8_79940130_C/T associated with


Abscission Rate. (C) chr6_16913820_G/C associated with Flower Amplitude. (D) chr7_633641_G/A associated with


Start Flowering DAS. For each locus, the left panel shows a regional association plot based on physical genomic


position for ETH23, ETH24 and the combined dataset, using four GWAS models: MLM, MLMM, FarmCPU and BLINK.


The dashed horizontal line indicates the significance threshold, and the red vertical line marks the focal SNP. For the


LD analysis, the focal SNP and its 12 nearest neighbouring SNPs on either side are shown. Gene annotations are


displayed below the regional association plot, with black arrows indicating transcriptional orientation and the gene


containing the focal SNP highlighted in red. The selected SNPs are connected to the triangular LD heatmap, which


displays pairwise linkage disequilibrium as 𝑟 [2] . The right panel shows the relationship between allele frequency at the


focal SNP and trait BLUEs for ETH23, ETH24 and the combined dataset. Coloured lines represent dataset-specific linear


fits, and the boxes report the fitted equation, the significance of the slope (𝑝) and the proportion of variance explained


(𝑅 [2] ).


[bioRxiv preprint doi: https://doi.org/10.64898/2026.11.03.763633; this version posted November 4, 2026. The copyright holder for this](https://doi.org/10.64898/2026.11.03.763633)

preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in
[perpetuity. It is made available under a CC-BY 4.0 International license.](http://creativecommons.org/licenses/by/4.0/)

## Tables


**Table 1** Description of the field experiments and data used for flowering dynamics modelling. For each experiment, the sowing date, the number of accessions (including spatial checks)and


plots planted in the field as well as the median number of plots per accession (excluding checks) are reported. Additionally, the maximum, median and minimum number of time points at

which images were taken per plot are depicted. The mean duration and standard deviation between the first and the last image taken is also shown. Finally, the number of plots successfully


modelled in terms of flowering and abscission phase are also provided.































**ETH23** 31.05.2023 169 4 594 3 15 22 30 95±19.3 591 559


**ETH24** 15.05.2024 169 4 564 3 20 31 35 121±21.8 564 557


**LfL23** 24.05.2023 169 7 215 1 7 10 13 81±15.1 215 207


**LfL24** 26.06.2024 169 1 358 2 7 9 14 70±10.9 358 354


**ETH24L** 09.08.2024 55 1 120 2 10 12 12 65±5.5 120 115


[bioRxiv preprint doi: https://doi.org/10.64898/2026.11.03.763633; this version posted November 4, 2026. The copyright holder for this](https://doi.org/10.64898/2026.11.03.763633)

preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in
[perpetuity. It is made available under a CC-BY 4.0 International license.](http://creativecommons.org/licenses/by/4.0/)


**Table 2.** Overview of marker-trait associations and their co-localization with annotated genes. The total number of


marker-trait associations, number of associations within a gene, unique number of genes associated to traits as well as


how many of them present an association in the coding sequence (CDS).













Image-derived

199 150 98 63
Flower Period


Image-derived

53 44 34 13
Flower Value


Image-derived

94 73 43 21
Abscission


Image-derived

343 265 171 95
all