Construction and Evaluation of Machine Learning Models for Near-Real-Time Fire Detection from MTG FCI Imagery

Asaf Vanunu; Boaz Nadler; Arnon Karnieli

cs.CVOct 4, 2026

Construction and Evaluation of Machine Learning Models for Near-Real-Time Fire Detection from MTG FCI Imagery

Authors: Asaf Vanunu, Boaz Nadler, Arnon Karnieli

Abstract

Geostationary satellite observations are important for wildfire detection and monitoring. The current study evaluates machine learning models for MTG FCI near-real-time fire detection in 1- and 2-km spatial configurations and compares them with threshold-based algorithms. The models were trained and evaluated using VIIRS fire reference data across diverse ecological regions in Europe, Africa, and the Middle East. The key results are that 1-km models significantly outperform both their 2-km variants and operational threshold products. The constructed 1-km models achieved F1 scores higher by up to 0.36 compared to baseline products. Importantly, the 1-km models detected small fires with higher probability compared to competing models. Finally, the models robustly detected fires up to 260 min earlier than baseline products. To support opensource applications, our trained models are publicly available.

Figures & tables

Figure 1: Study areas over land cover maps, covering Europe and Central Africa. The boundaries of training regions are delinted by black line, the evaluation regions by red, and the regions used for both are marked in light blue.

Region	Months	Number of MTG FCI Images	Number of VIIRS Fire Detections
Balkan	July, August, September 2025	159	8,387
Germany		138	11,125
Middle East		109	3,190
Spain - Portugal		143	15,197
Turkey		151	20,702
Nigeria	January, February, March 2026	155	42,275

Table 1: Summary of the evaluation dataset.

Band ( $\mu$ m)	HR (km)	NR (km)
VIS 0.6	0.5	1.0
NIR 2.2	0.5	1.0
MIR 3.8	1.0	2.0
TIR 10.5	1.0	2.0

Table 2: Selected MTG FCI spectral bands used as ML models input.

Figure 2: Locations of 20 fire events across Europe during summer 2025.

Resolution	Fire Pixels	Non-Fire Pixels
1-km (HR)	82,466	981,000
2-km (NR)	75,186	981,000

Table 3: Summary of the tabular training datasets.

Resolution	Fire Patches	Non-Fire Patches	Fire Pixels	Non-Fire Pixels (Million)
1-km (HR)	14,174	5,515	153,620	325
2-km (NR)	7,221	1,850	113,847	129

Table 4: Summary of the patch-based datasets.

Figure 3: Probability density distributions of $\mathit{BT}_{3.8\,\mu\text{m}}$ for TP fire detections compared between 1 and 2-km model variants of (a) CatBoost, (b) MAnet, (c) U-Net++, and (d) the FRP-PIXEL and FIR products. For each model, a two-sample Kolmogorov-Smirnov test confirmed a statistically significant difference between the 1-km and 2-km distributions ( $p<0.001$ ).

Figure 4: Scatter plots comparing pixel-matched $\mathit{BT}_{3.8\,\mu\text{m}}$ for TP fire detections between 1-km and 2-km variants of (a) CatBoost, (b) MAnet, and (c) U-Net++.

Figure 5: Overall recall, precision, and F1 scores aggregated over all evaluation regions for the 1-km and 2-km models alongside the FIR and FRP-PIXEL.

Figure 6: Comparison of fire detection outputs in Germany on 11 August 2025 (00:50 UTC). The panels display predictions from the 1-km MAnet, U-Net++, and CatBoost models, alongside the 1-km FRP-PIXEL and 2-km FIR. Remapped VIIRS fire detections (00:48 UTC) are provided as reference data.

Figure 7: Probability of fire pixel detection by ML, FIR, and FRP-PIXEL models as a function of the number of VIIRS fire detections ( $k$ ) inside an MTG FCI pixel.

Total Fire Events	20
	vs. FIR (2-km)			vs. FRP-PIXEL (1-km)
Metric	CatBoost	MAnet	U-Net++	CatBoost	MAnet	U-Net++
Earlier Detections ( $N_{\text{win}}$ )	12	13	13	4	8	8
Same Time Detections ( $N_{\text{same}}$ )	7	7	7	14	12	12
Missed / late Events ( $N_{\text{miss}}$ )	0	0	0	2	0	0
Probability of Win ( $P_{\text{win}}$ )	0.63	0.65	0.65	0.20	0.40	0.40

Table 5: Early detection metrics comparing 1-km models to the FIR and FRP-PIXEL. The metrics include total early detection counts (

N_{\text{win}}

), simultaneous detections (

N_{\text{same}}

), missed detections (

N_{\text{miss}}

), corresponding probabilities (

P_{\text{win}}

and

P_{\text{same}}

), and the conditional expected lead time (

\mathbb{E}[\Delta t\mid\text{win}]

) in minutes.

Figure 8: Time difference in initial fire detection between the 1-km ML models and (a) FIR, and (b) FRP-PIXEL

Figure 9: Sequence of detections for the Plasencia fire event. Illustrating early detection by the MAnet and U-Net++ models prior to CatBoost, FIR, and FRP-PIXEL.

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

Figure B.1: Ten most important features for the (a) 1-km (HR) model and (b) 2-km (NR) model.

Figure D.1: Comparison of recall, precision, and F1 scores across all evaluation regions for the 1 and 2-km ML models alongside the FIR (2-km) and FRP-PIXEL (1-km) benchmark products.

Figure E.1: Region-specific probability of fire detection as a function of the number of VIIRS fire detections per MTG FCI pixel ( $k$ ).

Explore similar work

Comparing a gradient boosting algorithm to the GOES FDC for wildfire detection
Oct 1, 2026Asaf Vanunu, Boaz Nadler, Arnon KarnieliWildfire Detection Xgboost Model
The Albert Katz International School for Desert Studies, The Jacob Blaustein Institutes for Desert Research, Sde Boker Campus, Ben-Gurion University of the Negev, Midreshet Ben-Gurion 84990, Israel · The Goldman Sonnenfeldt School of Sustainability and Climate Change, Ben-Gurion University of the Negev, Beer Sheva 8410501, Israel · The Remote Sensing Laboratory, French Associates Institute for Agriculture and Biotechnology of Drylands, The Jacob Blaustein Institutes for Desert Research, Sde Boker Campus, Ben-Gurion University of the Negev, 84990, Israel +1
On-Orbit Real-Time Wildfire Detection Under On-Board Constraints
May 7, 2026Matthias Rötzer, Veronika Pörtge, Martin Ickerott +6Wildfire Detection Rgb-Thermal
OroraTech GmbH, St.-Martin-Str. 112, 81669 Munich, Germany · Ludwig-Maximilians-Universität München, Munich, Germany
Set Prediction for Next-Day Active Fire Forecasting
May 11, 2026Yuchen Bai, Georgios Athanasiou, Xin Yu +4Wildfire Prediction
Max Planck Institute for Biogeochemistry, Jena, Germany · Orion Lab, National Technical University of Athens · University of Utah, Salt Lake City, USA +4

Oct 1, 2026, cs.CVJ/K move · Enter open · S save

Comparing a gradient boosting algorithm to the GOES FDC for wildfire detection

Asaf Vanunu, Boaz Nadler, Arnon Karnieli

The Albert Katz International School for Desert Studies, The Jacob Blaustein Institutes for Desert Research, Sde Boker Campus, Ben-Gurion University of the Negev, Midreshet Ben-Gurion 84990, Israel · The Goldman Sonnenfeldt School of Sustainability and Climate Change, Ben-Gurion University of the Negev, Beer Sheva 8410501, Israel · The Remote Sensing Laboratory, French Associates Institute for Agriculture and Biotechnology of Drylands, The Jacob Blaustein Institutes for Desert Research, Sde Boker Campus, Ben-Gurion University of the Negev, 84990, Israel+1

Wildfire Detection Xgboost Model Query Image Gradient

Open paper AskPDF arXiv

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.

Abstract

Explore similar work

Comparing a gradient boosting algorithm to the GOES FDC for wildfire detection

On-Orbit Real-Time Wildfire Detection Under On-Board Constraints

Set Prediction for Next-Day Active Fire Forecasting