Source-linked AI summary
Comparing a gradient boosting algorithm to the GOES FDC for wildfire detection
Asaf Vanunu, Boaz Nadler, Arnon Karnieli
TL;DR
Wildfire detection from geostationary imagery matters because fires threaten people, ecosystems, and property, while prior machine-learning studies left broad-area and operational comparisons unresolved. This study trains a CatBoost detector on VIIRS-labeled GOES ABI imagery and compares it with GOES FDC. CatBoost performs better across the five evaluated regions, including regions outside the training area.
Problem
Prior GEO fire-detection studies often used small datasets and regions, did not compare models with operational algorithms, and omitted separate day/night evaluation, leaving broad-area performance uncertain.
Source: §1 · p. 4Method
The study trains a CatBoost model on VIIRS-referenced GOES ABI imagery and evaluates it against GOES FDC, including regions and a time period not used for training.
Source: §1 · p. 5, §5 · p. 27Results
CatBoost achieved higher precision, recall, and F1-scores than GOES FDC across all five regions, with stronger performance also reported outside the training domain.
Source: §5 · p. 27Takeaways & Limitations
The study concludes that CatBoost can improve GOES ABI wildfire detection with higher accuracy, lower false-alarm rates, and earlier detection than GOES FDC.
Source: §5 · p. 29Testing covered continental U.S. and parts of North America using CONUS and PACUS scan modes, leaving full-disk and regions such as the Amazon or Patagonia unevaluated.
Source: §5 · p. 29
Abstract
from arXiv · showhide
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.
1. Introduction
Wildfires threaten people, ecosystems, and property, making timely detection important. The study develops a GOES ABI machine-learning detector and evaluates it against GOES FDC across regions, time periods, and day/night conditions.
Motivation
- Timely detection matters because wildfires threaten life, ecosystems, and property, and the frequency of major wildfire disasters quadrupled from 1980 to 2023.The introduction gives the 2025 Palisades Fire as an example of a severe event.Source: §1 · p. 3
Background
- Geostationary satellites monitor the same region every 5 to 15 minutes, while existing GEO fire products can miss fires visible in the 3.7 µm band and have high false-alarm rates.The cited concerns include GOES FDC and MSG FRP-PIXEL products.Source: §1 · p. 3
Machine-learning approach
- Machine-learning models may learn complex spatial and spectral fire patterns from GEO imagery, using pixel-level labels from independent sources and producing binary fire/non-fire images.Prior work has used Random Forest and CNN approaches, but important limitations and gaps remain.Source: §1 · p. 4
Research gap
- Prior GEO fire-detection studies often used small datasets and geographic areas, lacked comparisons with operational algorithms, and did not separately evaluate daytime and nighttime performance.These gaps left it unclear whether machine-learning models outperform existing algorithms across broad areas and atmospheric conditions.Source: §1 · pp. 4–5
Study objective
- The study trains a gradient-boosting fire detector on VIIRS-labeled GOES ABI imagery and compares it with GOES FDC using separate data from other regions and a later year.It also compares overall and daytime/nighttime detection performance.Source: §1 · p. 5
2. Study area and data sources
The study uses GOES ABI imagery and VIIRS fire detections to train and evaluate a fire detector across multiple regions, and compiles perimeter data for early-detection comparisons.
2.1. Study area
- Training and evaluation cover distinct and overlapping areas across the United States, Mexico, and Canada, with approximately 2.1 million km2 used for training and 2 million km2 for evaluation.The map marks training-only states in black, evaluation-only states in red, and states used for both in light blue.Source: §2.1 · pp. 5–6, Figure 1 · p. 6
2.2. Data sources
- Training uses all study-area VIIRS detections from 2022–2023: 324,502 detections matched to 4,330 GOES ABI images; evaluation uses selected months in 2024.The datasets pair VIIRS detections with ABI images acquired within ±2.5 minutes.Source: §2.2 · pp. 6–7
2.2.1. VIIRS active fire product
- VIIRS VNP14IMG Collection 2 provides the reference fire labels at 375 m resolution, with low-confidence detections excluded.The product also provides location, acquisition time, fire radiative power, and confidence level.Source: §2.2.1 · p. 7
2.2.2. GOES ABI images
- The 2-km MCMI and ACM products serve as model inputs, while FDC is the benchmark; early-detection tests also include FDC’s high category despite its higher false-alarm rate.ACM supports cloud masking, and MCMI contains 16 multispectral bands.Source: §2.2.2 · pp. 7–8
2.2.3. Fire perimeter data for early detection assessment
- Perimeter data for 51 fire events from WFIGS support early-detection comparisons, and Figure 2 maps the events with identification numbers.Fire number 19 is identified as the Bobcat Fire in Florida on 29 May 2021.Source: §2.2.3 · p. 8, Figure 2 · p. 9
3. Methodology
The methodology aligns VIIRS fire detections with GOES ABI imagery to build labeled pixels, then trains a CatBoost classifier using pixel values and neighborhood context. It also specifies the model-evaluation pipeline.
3. Methodology
- The workflow assigns VIIRS detections to the GOES ABI grid, constructs labeled data, trains CatBoost, and evaluates the resulting fire detector.The authors also provide the trained model's code and example data in a GitHub repository.Source: §3 · p. 9
3.1. Assigning VIIRS fire detections to GOES ABI grid
- VIIRS fire-detection centroids are rasterized onto the GOES ABI grid, and residual duplicate VIIRS pixels from overlapping scan lines are removed.Each ABI-grid pixel records the number of VIIRS centroids intersecting it; pixels without intersections receive 0.Source: §3.1 · pp. 9–10
3.2. Labeled dataset construction
- A 3×3 window around each rasterized VIIRS detection labels the GOES ABI pixel with the highest Fire Index as fire, accommodating spatial misalignment.For non-fire labels, the selected pixels must be outside 3×3 windows around labeled fire pixels and VIIRS pixels with more than one detection; contextual features exclude cloud-covered neighbors.Source: §3.2 · p. 11, §3.9 · pp. 11–12
3.9 µm and 11.2 µm bands, respectively. The 3.9 µm band is sensitive to fires,
- The Fire Index uses GOES ABI brightness temperatures at 3.9 µm and 11.2 µm; high values are associated with active fires.The 3.9 µm band is sensitive to fires, whereas the 11.2 µm band is not.Source: §3.2 · p. 11, §3.9 · p. 11
3.3. Model training
- CatBoost builds an ensemble of decision trees, with each tree correcting earlier trees' errors; training data were split 80% for training and 20% for validation.The method uses ordered boosting, native categorical-variable handling, and oblivious decision trees.Source: §3.3 · p. 13
Appendix B.
The evaluation matches GEO fire detections to rasterized VIIRS detections within spatial windows, then calculates recall, precision, and F1 from dataset-wide totals.
Appendix B.
- A GEO detection is a true positive when a rasterized VIIRS detection lies within its 3×3 window; otherwise it is a false positive, while unmatched VIIRS detections are false negatives.Figure 5 illustrates the GEO detections, VIIRS detections, and resulting classifications.Source: §3.4 · p. 14, Figure 5 · p. 14
- Recall measures the share of true fire pixels detected, precision the share of flagged pixels that are true fires, and F1 their harmonic mean.All three metrics range from 0 to 1, with higher values indicating improved accuracy.Source: §3.4 · p. 14
- Recall, precision, and F1 are calculated from the summed true-positive, false-positive, and false-negative counts across the image dataset.The method denotes the dataset size by n and records each count for image i before summing totals.Source: §3.4 · pp. 14–15
4. Results
The results compare CatBoost with GOES FDC across five regions, including regions outside the training area, for accuracy, day/night performance, early detection, training-data effects, and fire-size sensitivity.
4.1. CatBoost performance against the GOES FDC
- CatBoost achieved F1 scores 0.16–0.38 higher than GOES FDC across all five evaluation regions, with lower false-alarm rates and equal or higher detection rates.It also performed better in recall, precision, and F1 in California–Oregon, Arizona–Utah, and British Columbia, which were outside the training area.Source: §4.1 · p. 15, Figure 6 · p. 16
4.2. Daytime/nighttime performance
- At night, GOES FDC recall was 0 in Georgia and Texas and nearly 0 elsewhere, while CatBoost F1 ranged from 0.17 to 0.69 versus FDC’s 0.01 to 0.1.During daytime, CatBoost F1 was higher in every region; recall in California–Oregon and precision in Georgia and British Columbia were almost identical between methods.Source: §4.2 · pp. 17–18, Figure 8 · p. 19
4.3. Impact of training dataset size on CatBoost performance
- The 66,000-fire-pixel model had higher precision than both 2,000-pixel models in Texas, California–Oregon, and Arizona–Utah, while recall was generally similar.The April–May-only model’s Georgia recall was 0.21, versus 0.33 for the 66,000-pixel model and 0.35 for the 2,000-pixel model; the 66,000-pixel model also had the highest F1 in the two named evaluation-only regions.Source: §4.3 · pp. 19–20
4.4. Early detection capability
- Across 51 fire events, CatBoost detected 26 fires before both VIIRS and GOES FDC in the first analysis, compared with six first detections by FDC.In the second analysis, CatBoost detected 23 fires first and FDC detected nine; in the 38-event comparison, 22 events were first detected by CatBoost using processed and saturated FDC categories, versus six by FDC.Source: §4.4 · pp. 21–22, Figure 10 · p. 23
4.5. Fire detection probability
- CatBoost had higher fire-detection probabilities than GOES FDC in California–Oregon, Arizona–Utah, and British Columbia, and was more sensitive to small fires in most cases.In California–Oregon, at k = 1, the probabilities were 0.46 for CatBoost and 0.28 for GOES FDC; at k = 5, they were 0.78 and 0.48, respectively.Source: §4.5 · p. 25, Figure 11 · p. 26
5. Discussion and conclusions
The CatBoost model outperformed GOES FDC across five regions, including regions outside its training domain, and maintained reliable nighttime detection. The study also identifies limits in geographic scope and reference data.
5. Discussion and conclusions
- The study evaluates CatBoost against GOES FDC using VIIRS as the reference, focusing on accuracy, early detection, and fire-detection probability.It extends prior GOES-ABI work by evaluating larger geographic regions and comparing against the operational product.Source: §5 · pp. 27–28
- CatBoost consistently outperformed GOES FDC across five regions, with higher precision, recall, and F1 scores; it also generalized to three regions outside its training domain.Those regions had different climatic conditions and satellite view angles from the training data.Source: §5 · p. 27
- Reliable nighttime detection by CatBoost contrasted with GOES FDC nighttime recall near zero across all regions.ABI's coarse resolution relative to VIIRS also meant small VIIRS-detected fires were false negatives for both models.Source: §5 · p. 27
- The study's findings are consistent with prior GEO-based ML results, while its GOES-ABI evaluation adds operational benchmarking and regional-scale assessment.Kang et al. (2023) reported transferability to Australia, while Zhao and Ban (2022) evaluated four fires without comparison to GOES FDC.Source: §5 · p. 28
- The study was limited to North American regions and CONUS and PACUS scan modes; full-disk regions such as the Amazon and Patagonia remain untested.Only CatBoost was evaluated, and VIIRS was the sole reference source.Source: §5 · p. 29
Appendix A Fire events
Appendix A summarizes 51 historical fire events used in the first-time detection comparison.
Appendix A Fire events
- Each event's record includes the timestamp of its first VIIRS fire detection and its fire size in km2.The table also includes the fire name and state.Source: Appendix A · p. 35, Table A.1 · p. 36
- The table records each fire's centroid latitude and longitude alongside its identifying fire ID or name.These attributes characterize the events used in the comparison.Source: Appendix A · p. 35, Table A.1 · p. 36
- The table summarizes 51 historical fire events used for the first-time detection comparison.It lists fire name, first VIIRS detection timestamp, state, centroid coordinates, and fire size.Source: Table A.1 · p. 36
Appendix B CatBoost Hyperparameters
Appendix B describes CatBoost hyperparameters, the precision-optimized cross-validation search, and important features in the final model.
Appendix B CatBoost Hyperparameters
- Tree depth was set to 8, and L2 leaf regularization to 1; these settings control tree complexity and penalize large leaf outputs.The passage says the regularization reduces overfitting.Source: Appendix B · p. 37
- A bagging temperature of 5 adds randomness to training-sample weights, which the passage says also reduces overfitting.Source: Appendix B · p. 37
- A 5-fold cross-validation grid search optimized hyperparameters for precision to reduce false alarms.The described settings include a learning rate of 0.01 and 500 iterations.Source: Appendix B · p. 37
- The final model's important features include fire-index statistics, the 3.7µm band, and the 0.47µm visible band.Fire-index standard deviation incorporates neighboring-pixel information; the visible band may help distinguish day from night.Source: Appendix B · p. 37, Figure B.1 · p. 38