Comparing a gradient boosting algorithm to the GOES FDC for wildfire detection
Operational geostationary fire products like the GOES Fire Detection and Characterization (FDC) suffer from high false-alarm rates and poor nighttime recall. This study replaces the FDC's rule-based algorithm with a CatBoost gradient-boosting model trained on over 60,000 GOES ABI fire pixels labeled by VIIRS active-fire detections. By evaluating on large 2024 regions outside the training domain, the authors show that a comparatively simple ML model can substantially outperform the operational product in precision, recall, and early detection latency, especially after dark.
The paper makes a clear, well-supported empirical case that a CatBoost model trained on a large labeled dataset outperforms the operational GOES FDC in both accuracy and timeliness, though its conclusions are bounded by a single architecture and a reliance on VIIRS as the sole reference standard. As the authors state, "Overall, this study shows that a CatBoost-based model can improve wildfire detection from GOES ABI imagery. It provides higher accuracy, lower false-alarm rates, and earlier detection than the GOES FDC product."
The experimental design is a major strength: training uses 2022–2023 data from the southeastern and south-central U.S. and northern Mexico, while evaluation uses separate 2024 regions spanning the western U.S. and British Columbia (~2 million km$^2$). The nighttime performance gap is stark and unambiguous: "the FDC recall is 0 in Georgia and Texas, and it is almost 0 in the other regions. This means the FDC fails to detect almost all fires in the nighttime images of the study dataset." The early-detection and fire-size probability analyses (Sections 4.4–4.5) add practical value beyond aggregate $F1$ scores.
A central limitation is the circular reliance on VIIRS as the only ground-truth source for both training labels and evaluation metrics, which the authors acknowledge: "the evaluation relied exclusively on VIIRS as a reference. Other LEO satellites with higher spatial resolution or in situ fire datasets were not used." Additionally, the training set was explicitly filtered to exclude small fires that ABI struggles to detect—requiring $\geq 2$ VIIRS detections in a $3\times 3$ window and $FI > 0$—yet Section 4.5 still claims superior sensitivity to small fires. The study also tests only CatBoost and provides no ablation against other ML models (e.g., Random Forest, XGBoost, or CNNs), leaving unclear how much of the gain comes from the model versus the large training corpus.
The quantitative evidence strongly supports the headline claims. The authors report that "the CatBoost ML model achieved F1 scores that were 0.16-0.38 higher than those of the GOES FDC," and in 51 retrospective fire events the model detected 26 fires before both VIIRS and GOES FDC versus only six earlier detections by the FDC. Comparisons to related work are fair and accurate: they correctly position their contribution against Zhao and Ban (2022) by noting that prior GOES-ABI ML work used a small number of events and "did not include a comparison with the operational GOES Fire Detection and Characterization product."
Reproducibility is relatively strong for the domain. The authors share the trained model and example code on GitHub and deposit primary data in Zenodo. Key CatBoost hyperparameters are reported (learning rate 0.01, 500 iterations, depth 8, L2 leaf regularization 1), and the 5-fold CV search objective is disclosed. "Code of the trained ML model and some example data are available at the author’s GitHub repository," while "The primary data that support the findings of this study are openly available in Zenodo at https://doi.org/10.5281/zenodo.19047285." Nevertheless, full reproduction requires rebuilding a massive training corpus from VIIRS and GOES archives, and fine-grained details such as random seeds, exact validation indices, and the bow-tie duplicate-removal implementation are not specified in the text.
Wildfires pose severe risks to human life, ecosystems, and property. This study presents a machine learning approach for wildfire detection from GOES ABI imagery. A CatBoost model was trained on a large dataset with thousands of ABI images and over 300,000 matching VIIRS fire detections. An evaluation on a separate dataset across five regions showed that the learned CatBoost model outperformed the operational GOES Fire Detection and Characterization (FDC) product. It achieved higher precision, recall, and F1 scores both within and outside the training area. The CatBoost model achieved F1 scores that were 0.16 to 0.38 higher than the GOES FDC in all regions. In addition, out of 51 historical fire events, the CatBoost detected 26 fires before both VIIRS and GOES FDC, compared to only six earlier detections by the GOES FDC. Importantly, the CatBoost model achieved accurate wildfire detection also during nighttime, whereas the GOES FDC obtained very low recall values, around 0.03. This study demonstrates that machine learning models may offer significant improvements over existing geostationary fire products, including higher accuracy, fewer false alarms, and earlier detection.
Pick a starting point or write your own. Challenges run in the background, so you can keep reading while the AI investigates.
No challenges yet. Disagree with the review? Ask the AI to revisit a specific claim.