Satellite history improves next-week fire-detection rankings in Bolivia benchmark

What the model predicts

The study divides Bolivia into 37,146 eligible grid cells with edges of 0.05 degrees. Each Monday, the model estimates the probability that at least one selected fire detection will occur inside each cell during the following seven days. One cell during one forecast week is a cell-week.

The observations come from the Visible Infrared Imaging Radiometer Suite, or VIIRS, aboard the Suomi National Polar-orbiting Partnership satellite, distributed through NASA’s fire-information service. The primary analysis retains records classified as presumed vegetation fires with nominal or high detection confidence. Those confidence categories describe the satellite detections, not the model’s forecast probabilities.

The weekly dataset spans 2013–2025. Importantly, a week without a recorded detection is not automatically counted as a negative example. The benchmark requires complete target-week observation coverage and complete observation history over the preceding 28 days. Documented outages and missing acquisitions can therefore exclude a forecast from evaluation.

Learning from earlier detections

The main model uses LightGBM, a machine-learning method that combines decision trees. Its inputs describe recent detection counts, time since the last detection, reported fire radiative power, activity in neighboring cells and across Bolivia, and seasonal patterns. It also uses historical detection frequencies calculated from training data.

Two simpler approaches provide comparisons. Climatology estimates the usual detection frequency for a location and time of year. Persistence estimates probability from the number of detections in the preceding seven days.

Model selection used 2021 observations. A separate adjustment using 2022 observations, called calibration, mapped model scores to estimated probabilities. Primary model choices were fixed before scoring the 2023–2025 test years.

This was nevertheless a historical simulation, not a demonstration using inputs archived at each actual forecast time. It assumes that archived standard-processing satellite records approximate the information that would have been available then—an assumption the study could not verify.

Better rankings in later years

The primary time-based test evaluated 3,127,636 eligible cell-weeks outside the geographically reserved regions. Detections occurred in 1.95% of those examples.

The calibrated LightGBM model achieved average precision of 0.323, compared with 0.167 for climatology and 0.143 for persistence. Average precision measures how effectively a ranking concentrates detection-positive examples near its top; it is not the percentage of forecasts that were correct.

The model’s average-precision advantage over climatology was 0.1567, with a paired 95% uncertainty interval of 0.0998–0.2179. Because that interval remained above zero, the comparison met the study’s predefined improvement criterion. The uncertainty calculation repeatedly resampled calendar-month blocks rather than treating every cell-week as independent.

The area-budget results make the ranking more tangible. Reviewing the highest-ranked 5% of eligible area each week would have covered 38.4% of cell-weeks with detections. At a 10% area budget, coverage rose to 55.8%, compared with 39.7% for climatology. These are retrospective coverage calculations—not measured gains from deploying monitoring teams or response crews.

A separate geographic test

The study also reserved 24 geographic blocks whose detection outcomes were excluded from fitting, selection and calibration. A separately selected model, without cell-specific climatology, achieved average precision of 0.384 there, compared with 0.181 for persistence.

Those locations still supplied their own earlier satellite observations as inputs. The result therefore does not establish performance in places without detection histories. Nor does the higher geographic score show that geographic transfer was easier: the test populations, predictors and comparators differed.

Environmental additions remain unresolved

Adding retrospective rainfall information produced no clear improvement in average precision. Its paired difference interval crossed zero; this does not prove that rainfall and the original model are equivalent.

The primary validation comparisons for vegetation condition, land cover and their combination were also inconclusive. Some secondary analyses and exploratory test comparisons produced positive intervals, but they did not change the primary decisions. These predictive comparisons cannot establish causal effects of rainfall, vegetation or land use.

Limits before practical use

The model still overestimated detection frequency in its highest-probability group. Observation filters retained only 106 of 157 test issues, so the reported performance does not describe every week or missing-observation period.

Satellite detections are also an imperfect measure of burning. Clouds, smoke, overpass timing, resolution and fire intensity affect what is observed. Multiple detections can represent one fire, while a cell without a detection can still contain undetected burning. The results cannot be translated into fractions of individual fires, ignition forecasts, burned area or incident severity.

The uncertainty intervals rely on 28 observed test months and 24 held-out spatial blocks. They do not include model-selection uncertainty, sensor error or alternative boundary choices.

The paper calls for prospective evaluation: archive each week’s inputs and predictions before the target week, then measure ranking quality, probability reliability and usefulness at an agency’s actual review capacity. Historical benchmark scores alone do not establish operational warning value.

Source and status: This article is based on the supplied full paper, including its appendices—not only its abstract. The source is the preprint Predicting next-week VIIRS fire detections in Bolivia: a reproducible benchmark with temporal and geographic evaluation, arXiv:2611.08539v1.