Satellites and Machine Learning Sharpen Fire Risk Maps in Mountainous China

Forest fires are among the most destructive natural disturbances on Earth, and predicting when and where they will ignite depends on knowing how dry the fuel on the forest floor has become. A new study from Shanxi Province, on China’s Loess Plateau, shows that blending satellite soil-moisture measurements with reanalysis weather data and machine learning can substantially improve estimates of the world’s most widely used fire danger index, offering a blueprint for fire-risk mapping in mountainous terrain where weather stations are scarce and biased.

The research, published in Frontiers in Forests and Global Change, centers on the Canadian Forest Fire Weather Index (FWI) system, a physics-based framework used across the globe to translate routine weather observations into fire danger ratings. The system requires only four inputs—daily temperature, relative humidity, wind speed, and precipitation—but its inner workings are layered. Three fuel moisture codes track the wetness of dead organic material at increasing depths: the Fine Fuel Moisture Code (FFMC) responds to weather within two to three days, the Duff Moisture Code (DMC) integrates conditions over roughly twelve days, and the Drought Code (DC) reflects deep seasonal drought over about fifty-two days. From these codes, the system derives the Initial Spread Index, the Build-up Index, and finally the FWI itself, a composite measure of potential fire-line intensity.

Operational FWI calculations work well where stations are dense, but Shanxi poses a difficult test. More than eighty percent of the province is mountainous or hilly, with the Taihang Mountains to the east, the Lüliang Mountains to the west, and a chain of basins running down the center. Its 111 national automatic weather stations are mostly sited in accessible, open locations, leaving high-elevation slopes and forest interiors unobserved. Traditional interpolation methods such as Kriging tend to smooth away the very spatial heterogeneity that matters most for fire danger. Meanwhile, ERA5-Land, the high-resolution European reanalysis product, provides continuous gridded meteorology but carries known systematic biases in precipitation and near-surface winds over complex topography—biases that the FWI’s nonlinear equations can amplify during daily iteration.

The research team, based at the State Grid Shanxi Electric Power Company Electric Power Research Institute, built a station-supervised machine learning framework to correct these errors. They assembled 25,798 quality-controlled station-day records from March to November 2025, split chronologically so that the model was trained on spring and summer data and validated on autumn conditions it had never seen. Using XGBoost, a gradient-boosted tree algorithm prized for its interpretability, the team predicted the three moisture codes directly, then derived the remaining indices through the standard FWI equations—ensuring the machine learning respected the physics-defined structure of the system.

Crucially, the researchers ran a set of ablation experiments to isolate the value of each data source. The baseline scheme, M1, drove the FWI equations directly with ERA5-Land meteorology and struggled, explaining only about twenty-three percent of the variance in the final FWI. The second scheme, M2, used XGBoost to learn nonlinear corrections of the reanalysis biases, dramatically tightening the fit—for example, raising the coefficient of determination for the Drought Code from 0.524 to 0.761. The third scheme, M3, added fused passive-microwave soil moisture from the SMAP, AMSR2, and Chinese FY-3 satellite sensors, processed by the National Satellite Meteorological Center into a daily national product.

The addition of satellite soil moisture pushed performance higher across the board. The ERA5 plus soil moisture plus XGBoost configuration achieved coefficients of determination of 0.69 for FFMC, 0.63 for DMC, 0.77 for DC, and 0.43 for the overall FWI, while cutting the root-mean-square error for the ignition-sensitive FFMC from 18.74 to 15.15. Time-series analysis showed the fused model tracked sharp FWI spikes during autumn dry-wet swings that the raw reanalysis missed entirely, including an extreme event in early November that the baseline completely failed to capture. The soil-moisture features also suppressed the systematic underestimation of extreme high values that plagued both simpler schemes.

As a physical explanation, the microwave measurements make sense. Passive microwave sensors penetrate clouds and retrieve the dielectric constant of the topsoil, which is strongly coupled through energy and water exchange with the moisture of dead fuels lying on the forest floor. The team engineered multi-scale features from these retrievals—rolling 7-, 14-, and 30-day means capturing short-, medium-, and long-term moisture backgrounds, plus a high-temperature drought stress index combining temperature and soil water content—allowing the model to represent the soil-fuel system’s gradual drying and rewetting rhythms. They also added lagged predictors, cumulative precipitation sums, and sine-cosine encodings of the day of year to capture seasonal transitions, while taking explicit precautions against target leakage in the lagged variables.

Beyond model development, the study mapped where and when Shanxi’s fire danger peaks. The FFMC fluctuated violently with weather, regularly exceeding 80—the extreme risk range—during the low-humidity, windy spring, while the deeper codes accumulated drought steadily toward mid-August. The final FWI showed a single-peak annual structure, with late spring to early summer, from April through June, constituting the province’s absolute high-risk period. Statistically, FFMC and DC were nearly decoupled, with a correlation of just 0.134, demonstrating that surface fine-fuel flammability and deep soil drought are governed by different processes and must be modeled independently.

Geographically, the fire danger showed a clear south-high, north-low structure. The highest mean values of all four indices clustered in the central-southern and southwestern mountainous regions, around the edges of the Jinan Basin and the southern Lüliang range, roughly between latitudes 35 and 37 degrees north. The researchers attribute this to the southern region’s earlier spring warming, more intense solar radiation, and vigorous evaporation, amplified by terrain-driven winds on mountain slopes. Using K-means clustering on the standardized mean values of the four indices, they produced a five-level fire-weather zoning in which high-risk areas formed contiguous blocks in the south while the cooler, wetter north remained largely low risk—a pattern consistent with the province’s natural geographic conditions.

The authors are candid about limitations. No spatially independent validation or high-resolution dynamical baseline was performed, so extrapolation to unsampled mountain locations remains uncertain. The FWI system’s empirical constants were originally calibrated in Canadian boreal forests, and direct transfer to Shanxi’s warm-temperate deciduous and mixed forests may introduce seasonal biases. Passive microwave soil moisture represents only the near-surface layer, which limited the gain for the long-memory Drought Code and produced none for the DMC. Stations also under-sample the very microclimates where fires ignite, and the 0.01-degree working grid does not reflect the true resolution of the sensors. Still, the study demonstrates a practical path forward: correcting reanalysis products against ground truth, enriching them with satellite observations of the land surface, and quantifying exactly which data sources earn their place in the model. Future work will benchmark the approach against mesoscale atmospheric simulations and deep spatiotemporal architectures, and localize the FWI parameters for Chinese fuel types—steps toward fire danger maps precise enough to guide prevention at the scale of a single ridge or valley.

Subject of Research: Machine learning and satellite soil moisture fusion for improving fire weather index estimation in mountainous terrain

Article Title: Spatiotemporal variation characteristics and multi-combination modeling assessment of fire weather index in Shanxi Province

Article References: Lu, Z., Liang, J., Li, J., Yuan, H., Wang, S., & Duan, Z. (2026). Spatiotemporal variation characteristics and multi-combination modeling assessment of fire weather index in Shanxi Province. Frontiers in Forests and Global Change, 9, Article 1892417. https://doi.org/10.3389/ffgc.2026.1892417

Image Credits: AI Generated

DOI: 10.3389/ffgc.2026.1892417

Keywords: fire weather index, XGBoost, ERA5-Land, soil moisture, satellite remote sensing, Shanxi Province, forest fire risk, machine learning, Canadian FWI system, mountainous terrain, reanalysis bias correction, microwave remote sensing