Abstract
Introduction:
Accurate estimation of fire-weather indices in complex mountainous terrain is limited by biases in reanalysis meteorology and incomplete representation of surface-moisture heterogeneity.
Methods:
We developed a station-supervised framework that combines ground meteorological observations, ERA5-Land variables, terrain information, and microwave soil-moisture products from SMAP, AMSR2, and FY-3 sensors. XGBoost was used to estimate FFMC, DMC, and DC, after which ISI, BUI, and FWI were derived using the standard FWI equations. Three schemes were compared on a chronological validation set.
Results:
The ERA5 + SMC + XGBoost scheme achieved R² values of 0.69, 0.63, 0.77, and 0.43 for FFMC, DMC, DC, and FWI, respectively, and improved overall agreement relative to the ERA5-driven baseline. Study-period mean patterns indicated higher fire-weather potential in central-southern and southwestern mountainous areas, with the highest seasonal potential from spring to early summer.
Discussion:
Satellite-derived soil moisture improved regional agreement with station-derived targets for several FWI components, but spatial extrapolation to unsampled mountainous locations remains uncertain because no spatially independent or high-resolution dynamical validation was performed.
1 Introduction
Forest fires are one of the most significant natural disturbance factors in global forest ecosystems, not only causing immense losses of forest resources and destroying biodiversity, but also releasing large amounts of greenhouse gasses and aerosols into the atmosphere, profoundly impacting the global carbon cycle and climate change (Mai et al., 2025; Van Wilgen, 2025). In recent years, against the backdrop of a warming climate, the frequency and intensity of extreme high-temperature and drought events have continuously increased, leading to prolonged fire seasons in many regions globally and frequent extreme wildfire behaviors(Lou et al., 2026).
The Canadian Forest Fire Weather Index (FWI) system is currently one of the most widely used fire risk assessment systems worldwide (Chen et al., 2025). This system only requires four conventional meteorological parameters as input—temperature, relative humidity, wind speed, and precipitation—to systematically reflect the moisture content status of combustibles at different depths on the surface and fire behavior characteristics. However, Operational FWI calculation requires daily temperature, relative humidity, wind speed, and precipitation observations, together with previous-day moisture-code states (Van Wagner, 1987; De Jong et al., 2016). For regions with dramatic topographical changes and significant micro-meteorological characteristics, limited and unevenly distributed meteorological stations struggle to capture spatial heterogeneity of fire risk (Carvalho et al., 2011). Traditional spatial interpolation methods, such as Kriging interpolation and inverse distance weighting, often produce significant smoothing effects in mountainous areas with complex terrain, making it difficult to meet the demands of modern precise forest fire prevention and control.
With the development of satellite remote sensing and Earth system assimilation technologies, ERA5-Land reanalysis data and multi-source satellite remote sensing products provide new opportunities for large-scale, high spatiotemporal resolution fire risk assessment. ERA5-Land provides spatially continuous meteorological forcing data, but it is essentially a downscaled product based on coarse-resolution models, and there are systematic biases in simulating precipitation and near-surface wind speeds in complex terrain areas, which leads to differences between its directly calculated FWI and the real situation (Faal et al., 2025). At the same time, microwave remote sensing technologies, such as SMAP, AMSR2, and FY-3D microwave products, can penetrate clouds and directly obtain the dielectric constant of surface soil, thereby retrieving surface soil moisture content (Gao et al., 2022; Xie et al., 2023). A large number of ecological and forest fire behavior studies have shown that there is a strong coupled relationship of energy and water exchange between surface soil moisture content and dead fuel moisture content on the forest floor (Lewis et al., 2024).
Recent AI-based weather studies represent spatiotemporal dependence more explicitly than tabular learners. U-Net-type encoder–decoder networks have been used for radar-based precipitation nowcasting, while generative and physics-informed nowcasting systems can retain multiscale precipitation structures (Ayzel et al., 2020; Ravuri et al., 2021; Zhang et al., 2023). Graph-based architectures represent long-range spatial interactions in global weather fields (Lam et al., 2023), and selective state-space/Mamba models have recently been explored for spatiotemporal weather forecasting (Xiao et al., 2026). Satellite-derived soil moisture has also been examined in relation to FWI moisture codes (Atalay et al., 2024). However, limited work has quantified the incremental value of multi-source microwave soil-moisture covariates for station-supervised reconstruction of the internal FWI moisture codes. We therefore use XGBoost (Chen and Guestrin, 2016) as an interpretable tabular framework to isolate that covariate contribution, while treating direct benchmarking against deep spatiotemporal architectures as a distinct future evaluation.
Shanxi Province is located in the eastern part of the Loess Plateau, with mountainous and hilly areas accounting for over 80% of its total area, making it a typical arid and semi-arid continental monsoon climate. The province experiences less rain and more wind in spring, and with the advancement of ecological restoration projects such as the “Taihang Mountain Afforestation Project” in recent years, the fuel load in forest areas has sharply increased, leading to unprecedented pressure for forest fire prevention and control (Li et al., 2026b). Based on this, this study takes Shanxi Province as the research area, aiming to: (1) reveal the spatiotemporal evolution characteristics of FWI and related indicators in Shanxi Province based on high-quality ground meteorological station data; (2) construct a machine learning fusion framework integrating ERA5-Land and multi-source microwave remote sensing soil moisture; (3) evaluate the contribution of different data sources in fire risk modeling through systematic ablation experiments. The results of this study will provide theoretical basis and technical pathways for the development of high-precision fire risk products in complex terrain.
2 Materials and methods
2.1 Study area
Shanxi Province (34°34’—40°44’N, 110°14’—114°33’E) is located in North China, in the middle reaches of the Yellow River, with a total area of approximately 156,700 square kilometers (Figure 1). Topographically, Shanxi Province exhibits a typical landscape, with the Taihang Mountains in the east, the Lüliang Mountains in the west, and a series of bead-like basins in the center. There are significant differences in altitude, and micro-topographic climate characteristics are evident (Bao and Liu, 2025). The climate is a temperate continental monsoon climate, with an annual precipitation of approximately 400–650 mm, highly concentrated in summer (July–August). In spring (March–May), temperatures rise rapidly, relative humidity is extremely low, and strong winds are common, making it the most critical forest fire season in Shanxi Province (Pan et al., 2016). In terms of vegetation cover, mountainous areas are mainly dominated by Pinus tabulaeformis, Larix principis-rupprechtii, Platycladus orientalis, and secondary shrub forests, with thick accumulation of surface litter, posing a high fire risk (Wei et al., 2017).
Figure 1
2.2 Data
2.2.1 Meteorological station data
The high-precision ground meteorological observation data used in this study come from 111 national automatic meteorological stations within and surrounding Shanxi Province. As shown in Figure 1, the 111 stations provide broad regional coverage across Shanxi Province and its surroundings. However, conventional station siting is concentrated in accessible and relatively open locations; the network therefore cannot be assumed to fully resolve high-elevation, valley–slope, or forest-interior microclimates. We account for this limitation when interpreting station-supervised gridded outputs over complex terrain. The station observation database spans from 1 January 2025 to 4 March 2026. For the temporal analysis and model validation reported here, quality-controlled station-day records from 2 March to 14 November 2025 were used (25,798 samples from 111 stations). A chronological split was applied at the station-day record level: 20,638 samples from 2 March–25 September 2025 were used for training, and the remaining 5,160 samples from 26 September–14 November 2025 were used for validation. Station longitude and latitude provide the coordinate reference for matching station-derived FWI reference values with gridded ERA5-Land and fused soil-moisture covariates on the common working grid.
The operation of the Canadian Fire Weather Index (FWI) system has clear and strict physical requirements for input parameters. In conjunction with the data dictionary explanation of the original meteorological stations, this study extracted the following four core fields to construct the daily baseline derivation of FWI station-derived FWI reference values: temperature, relative humidity, wind speed, and precipitation (Atalay et al., 2024). Temperature directly determines the thermodynamic state of the atmospheric environment and is the core thermal factor driving the evaporation of surface litter and deep humus layer moisture, breaking the balance of effective precipitation interception. Relative humidity governs the moisture exchange gradient between fine dead fuels and the atmospheric boundary layer and is the most critical meteorological variable determining whether fuels are extremely dry. In fire risk modeling, wind speed not only accelerates convective evaporation from the surface but also controls the Initial Spread Index (ISI) and determines the key fluid dynamic parameter for the spread intensity of forest fires once they occur (Shang et al., 2024). Precipitation is the only liquid water input item in the FWI system and can directly reset or alleviate the accumulated drought status of various moisture codes (FFMC, DMC, DC).
2.2.2 Fire-weather indices
The FWI system is one of the most widely used forest fire warning and assessment frameworks globally. This system is essentially a physics-based, multi-level cumulative feedback model driven by meteorological conditions (Jiang et al., 2024). By inputting daily routine meteorological data, it systematically simulates the dynamic changes in moisture content of dead fuels at different depths within the forest, from the surface inwards, and then derives potential forest fire behavior characteristics (Wepryk et al., 2025). The entire FWI system has a rigorous internal logic from bottom to top, mainly divided into three fuel moisture codes and three fire behavior indicators (Van Vliet et al., 2024). The three underlying moisture codes, respectively, represent the dry-wet conditions of organic layers at different depths within the forest profile (FFMC, DMC, DC), with their core difference lying in the time lag of response to meteorological forcing. Based on these underlying moisture codes, the system further derives three comprehensive fire behavior indicators (ISI, BUI, FWI) that characterize the macroscopic dynamics and thermodynamic features of the fire scene (Chen et al., 2025).
Fine Fuel Moisture Code (FFMC): Reflects the moisture content of fine dead fuels, such as undecomposed litter, fallen leaves, and fine grasses on the surface. Because it is directly exposed to the atmospheric boundary layer, has a small water holding capacity, and undergoes intense moisture exchange, it is extremely sensitive to short-term meteorological changes, especially sudden drops in relative humidity and small amounts of precipitation. Its response time is only 2–3 days, making it the most active fire risk parameter and the most direct indicator for determining whether a forest fire can be ignited (Li, 2025).
Duff Moisture Code (DMC): Represents the drying degree of the duff layer, which is 5–10 cm below the surface and in a moderately decomposed state. Moisture exchange in this layer is relatively slow, with a response time of approximately 12 days. DMC is mainly affected by temperature and cumulative precipitation effects, and it can filter out short-term weather fluctuations, serving as an environmental background supplement to FFMC, reflecting whether a region is in a continuously dry state for 1 to 2 weeks (Gabban et al., 2008).
Drought Code (DC): Characterizes the long-term drought degree of the compact deep organic layer or underlying soil at depths of 10–20 cm and below. Its response time is as long as 52 days, it is extremely insensitive to single rainfall events, and it primarily reflects seasonal moisture deficit states. Extremely high DC values often mean that deep fuels are extremely dry, and once a forest fire occurs, it not only releases a huge amount of heat but also easily forms deep surface fires or underground fires that are difficult to extinguish (Di Giuseppe et al., 2018).
Initial Spread Index (ISI): Calculated by the nonlinear coupling of FFMC and near-surface wind speed. It combines extremely dry fine fuels with the horizontal driving force of the wind field, characterizing the initial spread rate of a fire after ignition, before being limited by large-scale terrain and complex fuel loads.
Build-up Index (BUI): Calculated comprehensively from DMC and DC. It represents the total background of dry organic matter available for combustion in the forest, reflecting the fuel reserve foundation that, if no effective precipitation occurs, the current forest can sustain a large fire and release heat (Van Wagner, 1987).
Fire Weather Index (FWI): As the final dimensionless output of the entire multi-level framework, FWI is comprehensively derived from ISI, which represents the spread potential, and BUI, which represents the background of heat release. Physically, it represents the fire line intensity under assumed standard flat forest conditions and is the core scale for comprehensively measuring the probability of forest fire occurrence, spread potential, and difficulty of suppression in a region. In summary, a higher FWI value means more intense energy release from the fire scene and a corresponding increase in fire risk level (De Jong et al., 2016).
2.2.3 Remote sensing data
ERA5-Land reanalysis data were used as the main spatially continuous meteorological forcing dataset in this study. ERA5-Land provides high-resolution land-surface atmospheric variables, including near-surface air temperature, precipitation, wind speed, and relative humidity-related variables, which are essential inputs for the Canadian Fire Weather Index system. Compared with sparse ground meteorological stations, ERA5-Land can provide continuous gridded meteorological information and is therefore suitable for regional fire-weather modeling in complex mountainous terrain. However, because reanalysis data may contain systematic biases in precipitation and near-surface wind fields, especially over heterogeneous topography, additional correction and data-fusion procedures are required.
The soil-moisture covariate used in this study was taken from the operational multi-source microwave soil-moisture fusion product documented by the National Satellite Meteorological Center in July 2024. At the upstream product-generation stage, FY-3B/FY-3C/FY-3D/FY-3G/FY-3F MWRI and AMSR2 L1 brightness-temperature observations are calibrated and fused using SMAP soil-moisture products as the reference. The dataset introduction states that SMAP has an original spatial resolution of 36 km and is processed to a 0.01-degree grid, FY-3/MWRI has an original spatial resolution of 25 km and is processed to 0.01-degree ascending and descending grids, and AMSR2 is provided on 0.01-degree ascending and descending grids. The operational product provides day and night retrievals from 1 January 2014 onward, with 0.01-degree global coverage. The fused soil moisture information was used as an important environmental constraint for the machine-learning inversion of FFMC, DMC, DC, and FWI, helping to capture the coupling relationship between atmospheric drying, soil water availability, and fuel moisture dynamics.
2.3 Data analysis
2.3.1 Data preprocessing
To construct a high-precision multi-source fusion inversion model, the primary task is to perform strict spatiotemporal alignment and quality control on heterogeneous data from different observation platforms. At the spatial scale, all raster products were resampled to a common 0.01° longitude–latitude working grid solely for feature collocation and output. This operation harmonizes grid geometry; it does not change the native or effective resolution of ERA5-Land or the passive-microwave products, nor does it add independent fine-scale atmospheric information. At the temporal scale, all observation and remote sensing data were extracted and aggregated to a daily scale. The core underlying parameters of the FWI system are essentially daily iterative physical models, meaning that the moisture code for a specific area on a given day must be calculated jointly from the previous day’s moisture code status and the current day’s meteorological forcing conditions. Therefore, daily temporal continuity is fundamental for model establishment. The 0.01-degree grid used in this study is a working grid for feature alignment and output, not the native or effective spatial resolution of the passive microwave sensors or ERA5-Land. Spatial patterns in gridded figures should therefore be interpreted as resampled and fused regional patterns rather than independent 0.01-degree sensor observations.
For the long-term original observation data obtained from national meteorological stations in Shanxi Province, this study developed a standard quality control scheme to minimize noise interference caused by instrument failures or transmission errors. First, data deduplication and temporal integrity review were performed to systematically remove redundant or duplicate observation records. For site time series with continuous missing data exceeding 3 days, to prevent human interpolation errors from being infinitely amplified in the daily iteration of FWI, they were treated as independent segmented sequences. For occasional single-day or short-term data gaps, smooth interpolation was performed by combining the spatial correlation characteristics of adjacent stations or by using the average observed values of the adjacent days before and after that station (Atalay et al., 2024). Second, outlier filtering and physical extreme value constraint identification were performed to remove outliers, while strict natural physical boundary conditions were superimposed for logical validation. Finally, to meet the requirements of the FWI system’s meteorological input, the 10-min average wind speed recorded by original stations was uniformly converted from m/s to km/h, temperature was unified to Celsius, and precipitation was kept in millimeters. After these rigorous preprocessing steps, a high-quality meteorological forcing basic dataset was ultimately constructed, featuring spatiotemporal continuity, strict physical logic, and standardized units, providing reliable data assurance for subsequent model training and true value baseline derivation.
2.3.2 Overall workflow
The overall technical roadmap of this study is shown in Figure 2, establishing a spatiotemporal modeling framework that combines statistical learning capabilities with FWI equation derivation. The entire process is based on the deep fusion of multi-source heterogeneous data, systematically integrating daily observation data from ground meteorological stations, soil moisture products inverted from multi-source passive microwave remote sensing, and ERA5-Land reanalysis data. Multi-scale data preprocessing is conducted under a unified data coordinate system (Chan et al., 2016; Silva et al., 2022). This specifically includes daily temporal scale alignment, 0.01° spatial resolution resampling, outlier identification and removal, and multi-variable unit standardization, thereby achieving strict consistency of data from different sources in temporal and spatial dimensions, providing a high-quality input foundation for subsequent modeling.
Figure 2
After completing the spatiotemporal coordination of basic data, the research design follows two mutually supportive, logically closed key technical paths. First, driven by high-precision measured meteorological elements from sites and relying on the classic physical mechanism of the Canadian Forest Fire Weather Index System, the process of fine fuel moisture content changes is calculated iteratively daily to obtain the station-derived FWI reference values of moisture codes, thereby constructing a supervised signal system with clear physical meaning. This workflow ensures that the learning targets are organized within the physics-defined FWI equation structure: FFMC, DMC, and DC are predicted first, and then ISI, BUI, and FWI are derived using the standard FWI equations. Additionally, multi-source environmental driving factors are systematically extracted and integrated, including surface moisture information sensitive to microwave remote sensing, meteorological elements from reanalysis data, and static constraint variables such as terrain factors. Through multi-scale feature representation and collaborative fusion, a high-dimensional feature space capable of comprehensively characterizing the moisture state of forest fuels and its driving mechanisms is constructed (Faal et al., 2025; Gao et al., 2022; Xie et al., 2023).
The study introduced various feature combination schemes to conduct systematic ablation experiments, quantifying the marginal contribution of different data sources and variable combinations to model performance, thereby identifying the optimal feature configuration scheme.
2.3.3 Calculation of the Canadian FWI system
When constructing the station-derived FWI reference values for the fire weather index, this study strictly followed the classic physical–mathematical model proposed by Van Wagner (1987), using a Markov chain process for daily iterative calculations. Each moisture code in the system is not only controlled by the current day’s meteorological forcing, namely precipitation interception and evapotranspiration, but also inherits the previous day’s fuel dryness state. The specific mathematical derivation process for each indicator is as follows:
(1) Fine fuel moisture code
The calculation of FFMC is divided into two stages: precipitation interception and natural temperature-humidity regulation. First, the previous day’s FFMC value () is converted into the initial moisture content. The calculations of the FWI components followed Equations 1–17.
When effective precipitation ( mm) occurs, the effective rainfall is set to , and is corrected for wetting using a specific exponential decay model:
The corrected is limited by the maximum water-holding capacity (250.0%). Subsequently, the drying/wetting equilibrium stage is entered. The drying equilibrium moisture content () and wetting equilibrium moisture content () are calculated:
Based on the relative magnitudes of with and , either the logarithmic drying rate or the logarithmic wetting rate , is chosen to calculate the final moisture content for the day. Taking the drying phase () as an example:
Finally, the moisture content is inversely converted to the FFMC value for the day (ranging from 0 to 101):
(2) Duff moisture code
DMC sets an effective precipitation threshold of 1.5 mm (). If effective precipitation occurs, the previous day’s DMC value () is converted to a corrected value after precipitation, based on the precipitation interception coefficient (which takes piecewise values depending on ). In rainless or post-rain conditions, the daily drying rate is driven by temperature and the characteristic day length coefficient () for the current month (Gu and Kang, 2025), The DMC value for the day () is calculated as:
(3) Drought code
DC represents deep soil moisture content, with an effective precipitation threshold of 2.8 mm (). After precipitation correction to obtain , the moisture consumption phase () is solely controlled by extreme temperature and the monthly potential evapotranspiration coefficient ():
(4) Comprehensive fire behavior indicators
After obtaining the three underlying moisture codes, the top-level fire behavior indicators are solved through nonlinear coupling. The Initial Spread Index (ISI) is determined jointly by the exponential amplification effect of wind speed and the decay function of FFMC moisture content:
The Build-up Index (BUI) comprehensively calculates DMC () and DC (). To prevent computational singularities, a double zero-value protection is set, and the equation is switched by boundary conditions:
The final comprehensive Fire Weather Index (FWI) uses a BUI-derived resistance/amplification factor () to scalar-correct ISI (intermediate variable ), and then outputs through exponential smoothing:
2.3.4 Quality control
After quality control, the daily station dataset contained 28,744 station-day records from 111 stations during 1 March–14 November 2025. Five station-day records were removed because relative humidity was below the valid threshold. For FWI derivation, station observations were aggregated to daily values using the 14:00 CST observation.
After applying the above QC rules and temporal matching, we constructed a cleaned modeling table containing 25,798 complete station-day samples. Samples were required to have valid targets (station-derived FFMC/DMC/DC/FWI reference values) and a complete set of predictor features after QC; otherwise, the corresponding station-day record was excluded.
To avoid target leakage, lagged FWI-code predictors used in M2 and M3 are not derived from station-observed target values; instead, they come from the ERA5-driven M1 FWI calculation using only previous-day information (m1_FFMC_lag1, m1_DMC_lag1, m1_DC_lag1, m1_FWI_lag1). Station-derived FFMC/DMC/DC/FWI values are used only as supervised targets in the training set and as reference values in validation.
2.4 Model evaluation and comparison
2.4.1 Correlation analysis and feature importance
For model input, this study used the final China daily fused soil-moisture product rather than reconstructing the upstream satellite fusion from raw brightness-temperature data. The upstream fusion is a DNN-based nonlinear correction and fusion workflow calibrated against SMAP, so no fixed sensor-weighted averaging was applied in this study. The daily China product was aligned with station and ERA5-Land covariates on the common 0.01-degree working grid used for feature extraction. This working grid is for model alignment and should not be interpreted as the native effective resolution of the individual microwave sensors. The dataset provides raster geometry, and each station day SMC value is extracted by nearest-neighbor matching from the daily grid using station longitude and latitude. Soil-moisture features in M3 include SMC, 7-, 14-, and 30-day rolling means. ERA5-Land variables are likewise matched to stations by nearest neighbor and then unit converted before bias correction.
Based on the completed feature space construction, to quantitatively evaluate the gain contribution of multi-source microwave remote sensing information in forest fire risk inversion under complex terrain conditions, this paper further designed three sets of progressively related ablation experiment frameworks based on the extreme gradient boosting tree, a typical nonlinear ensemble learning algorithm. Specifically: Scheme one (M1) was constructed as a traditional ERA5-only physical baseline model, directly calling reanalysis meteorological variables to drive the FWI physical iteration system, used to characterize the original meteorological forcing errors introduced by reanalysis data under complex underlying surface conditions (Silva et al., 2022); Scheme two (M2) was constructed as an ERA5-only machine learning model, taking the moisture codes inverted from meteorological stations as the supervised target and ERA5 meteorological variables as input features, performing nonlinear correction of meteorological biases through a data-driven approach to evaluate the correction capability of machine learning for reanalysis data errors (Emick et al., 2023); Scheme three (M3) further introduced multi-source microwave remote sensing soil moisture as a key environmental constraint variable based on M2, constructing a fusion-driven core inversion model to explore the actual contribution of microwave observations to characterizing surface moisture dynamics in fire risk modeling (Ishitobi et al., 2023).
2.4.2 Feature engineering parameter construction in the model building process
It is particularly worth noting that in the specific implementation of models M2 and M3, a multi-scale feature engineering system was systematically constructed to enhance the models’ ability to characterize the dynamic changes in fuel moisture. Specifically, during the feature construction phase, first-order time lags (lag-1) were built for ERA5 bias-corrected meteorological variables and the underlying moisture codes of the FWI system, based on site-by-site time series. This was done to represent the memory effect of fuel moisture status on previous meteorological conditions and its own state, thereby characterizing typical temporal inertia features. Second, multi-temporal scale rolling statistics were introduced to characterize cumulative meteorological influence processes, including 7-day and 14-day cumulative precipitation (pre_roll7, pre_roll14) and 7-day rolling average temperature (tmp_roll7), used to simulate the gradual response mechanism of the soil-fuel system during continuous wetting or drying processes (Jia et al., 2026).
To avoid target leakage, the lagged FWI-code predictors used in M2 and M3 are not lagged station-observed target values. They are one-day lagged values from the ERA5-driven M1 FWI calculation (m1_FFMC_lag1, m1_DMC_lag1, m1_DC_lag1, and m1_FWI_lag1), computed sequentially within each station time series using only previous-day information. The station-derived FFMC, DMC, DC, and FWI values are used only as supervised targets in the training set and as reference values in validation.
In terms of characterizing periodic features, sine and cosine transformations (doy_sin, doy_cos) were applied to the day-of-year to achieve continuous representation of seasonal changes, thereby avoiding the discontinuity problems caused by traditional discrete monthly variables and improving the model’s fitting ability for seasonal dry-wet transition processes.
Regarding the critical regulatory role of soil moisture on surface fuel moisture, multi-scale soil moisture features were further constructed, including original volumetric soil moisture content (SMC_vol) and its 7-day, 14-day, and 30-day rolling averages (SMC_vol_roll7, SMC_vol_roll14, SMC_vol_roll30), corresponding to short-term, medium-term, and long-term moisture background states, respectively, used to characterize the continuous supply capacity of soil moisture. Additionally, to strengthen the coupling relationship between air temperature and soil moisture, a high-temperature drought stress index (SMC_dry_temp) was constructed. Through a nonlinear combination of air temperature and soil moisture content, this index highlights the risk of extreme dryness under conditions of high temperature and low moisture content, thereby enhancing the model’s response capability to fire-sensitive environments (Miller et al., 2025).
Through the aforementioned multi-dimensional, multi-scale, and physically meaningful feature engineering design, the model can not only effectively capture the comprehensive impact of meteorological drivers, soil moisture status, and their interactions on fuel moisture, but also improve its ability to characterize the fire risk evolution process under complex terrain conditions, providing a solid data foundation for subsequent machine learning inversion (Ishitobi et al., 2023).
2.4.3 Comparative evaluation method
To determine the optimal structure of the model, this paper first used a grid search method to systematically optimize key hyperparameters, and the obtained optimal parameter combination was used as the training configuration for the final model. For model performance evaluation, the coefficient of determination (R2, Equation 18), mean absolute error (MAE, Equation 19), and root mean square error (RMSE, Equation 20) were selected as evaluation metrics to comprehensively measure the model’s predictive ability from two dimensions: goodness of fit and error distribution. This indicator system can effectively characterize the accuracy and stability of the model on different datasets, providing a quantitative basis for comparative analysis and subsequent optimization between models. The calculation formulas for each evaluation indicator are as follows:Where is the station-derived reference value for sample i, is the corresponding model prediction, M̅ is the mean of the n reference values, and n is the number of validation samples.
3 Results
3.1 Temporal variation characteristics of fire-weather indices
Based on daily-scale observations and calculation results from meteorological stations in Shanxi Province, Figure 3 reveals that fire weather indicators in Shanxi Province exhibit distinct and differentiated seasonal evolution patterns. As the most sensitive indicator to meteorological elements, the time series of the Fine Fuel Moisture Code (FFMC) (Figure 3c) shows characteristics of high-frequency violent fluctuations and extreme sensitivity to precipitation. From spring to early summer, due to extremely low relative humidity and frequent strong winds (Figure 3b), FFMC consistently remains in the extremely high-risk range above 80. Although intermittent precipitation in summer can cause it to drop sharply within 1–2 days, it quickly rebounds to above the flammable threshold under the drive of strong post-rain radiation. Unlike the high-frequency oscillations of FFMC, DMC and DC, which represent the drought degree of medium and deep organic matter layers, show obvious moisture accumulation and smoothing effects (Figure 3d). DMC shows a monotonically increasing trend from spring, reaching its annual peak in June–July, and then rapidly declines due to concentrated summer rainfall. DC, with the longest response lag, exhibits a cross-seasonal long-term drought accumulation characteristic. Its moisture deficit continuously accumulates from spring until it reaches its lowest point in mid-August, profoundly reflecting the extremely dry background of deep soil after prolonged evapotranspiration. Driven by this comprehensive underlying moisture code, the final Fire Weather Index (FWI) time series (Figure 3e) shows a highly concentrated single-peak jump structure throughout the year. That is, late spring to early summer (April to June) is the absolute extremely high fire risk period for the entire province, with frequent sharp pulses in FWI mean values during this period. After entering summer and autumn, with the arrival of the rainy season and falling temperatures, the fire risk potential rapidly decays to a safe baseline.
Figure 3
This difference in temporal evolution due to varying moisture response lags is further confirmed in the statistical distribution and correlation of each indicator (Figure 4). Probability density curves show that FFMC exhibits extremely strong dynamic changes (Figure 4a), with a median of 68.0 and the highest density interval approaching 90, indicating that surface fine fuels are in a flammable dry state most of the time. In contrast, DMC and DC are at a lower risk base for most periods (Figures 4b,c). Furthermore, scatter correlation analysis clearly indicates a significant statistical decoupling phenomenon between FFMC, which is dominated by short-term micro-meteorology, and DC, which is controlled by the long-term climate background (Figure 4e; R = 0.134). This physically proves that tiny surface wet-dry fluctuations and deep soil drought are not interdependent, further establishing the scientific necessity of independently inverting moisture codes at different depths in machine learning modeling.
Figure 4
Regarding the spatial distribution pattern, the inversion results selected for typical high fire risk periods (Figure 5) clearly reveal that the fire risk distribution in Shanxi Province exhibits a strong latitudinal gradient and terrain dependence. Whether it is surface FFMC, medium-deep DMC and DC, or comprehensively assessed FWI, their high-value core areas (dark red patches in the figure) are highly concentrated in the central-south and southwestern regions of Shanxi Province (roughly located in the northern latitude 35°-37° range, such as the edge of the Jinan Basin and the southern section of the Lvliang Mountains). The average FWI in this region is higher than in the northern high-latitude areas. This spatial heterogeneity is mainly attributed to the early convergence of cold and warm air masses in spring in the southern region, rapid temperature rise, and extremely vigorous surface evaporation. When superimposed with the dynamic driving force of the wind field on local mountain windward slopes, this forms spatial high-risk areas extremely prone to triggering large-scale forest fire spread.
Figure 5
3.2 Performance comparison of different combination schemes
By matching the independent validation set raster values output by the three models one-to-one with the station-derived FWI reference values from meteorological stations, Figure 6 intuitively displays the scatter distribution characteristics of each scheme. The traditional ERA5 direct physical calculation model (M1, orange point cloud) exhibits a significant discrete state with poor linear fit. Especially in the FWI scatter plot representing overall fire risk intensity (Figure 6j), its R2 is only 0.233, and it shows a severe systematic underestimation in the inversion of FFMC. When the XGBoost algorithm is introduced to construct the M2 model, the algorithm’s nonlinear correction capability for reanalysis meteorological biases begins to emerge. The point clouds for each indicator clearly converge toward the 1:1 diagonal line, for example, the R2 of DC jumps from 0.524 to 0.761. The M3 fusion model proposed in this study shows the optimal spatial fitting ability, with its scatter shape being the most compact. By deeply integrating microwave remote sensing soil moisture features, M3 not only further increases the R2 of the comprehensive FWI to 0.431 but also effectively suppresses the underestimation trend of M1 and M2 in extreme high-value areas.
Figure 6
In addition to spatial scatter assessment, capturing the dynamic change rhythm of moisture codes is equally crucial for early forest fire warning. Figure 7 selects the time series average results for a station during a continuous dry-wet alternating window period in autumn 2025. Station-derived FWI reference values show that meteorological conditions changed drastically during this period, triggering multiple steep FWI peaks. Due to the smoothing effect of coarse-resolution reanalysis data on local micro-meteorological forcing, the time series curve of M1 appears too sluggish, not only showing a noticeable delay in the rebound speed during the natural drying recovery period after rainfall but also completely missing the extreme FWI peak that occurred in early November. In contrast, the dynamic tracking capabilities of M2 and M3 are enhanced. Particularly for M3, by incorporating daily updated remote sensing surface soil moisture as an additional physically meaningful input related to surface dry–wet conditions, the model can more accurately capture the rapid fluctuations of observed values. Whether in the rapid daily fluctuations of FFMC or the slow accumulation process of deep DC, M3 demonstrates the best resilience in capturing extreme values during high forest fire risk periods.
Figure 7
Figure 8 further quantifies the error distribution of the three schemes from a statistical perspective. Overall, the M3 model, which integrates multi-source features, has the highest fitting accuracy. In terms of correlation (Figure 8a), M3 achieved the highest R2 for all four indicators; in terms of absolute error (Figures 8b,c), both the RMSE and MAE of M3 dropped to the lowest levels. Taking FFMC, which plays a decisive role in the initial ignition probability, as an example, the RMSE of M1 was as high as 18.74, while M3 reduced it to 15.15. This statistical result confirms that in non-flat forest areas with complex terrain, introducing microwave remote sensing soil moisture as a core feature can substantially compensate for the spatiotemporal uncertainties of traditional reanalysis meteorological data in calculating fire risk indices.
Figure 8
3.3 Spatial pattern of study-period mean fire-weather conditions in Shanxi Province
The average results obtained from the optimal inversion scheme, as shown in Figure 9, indicate that FFMC, DMC, DC, and FWI in Shanxi Province all exhibit significant spatial heterogeneity, but the spatial differentiation intensity varies among different indicators. Overall, the spatial variation range of the study-period mean FFMC is relatively small, generally maintaining a high background level, suggesting that surface fine fuels in Shanxi Province are generally in a relatively easy-to-dry state on an annual average. Its high-value areas are mainly distributed in the central-southern mountainous belt and local southwestern areas, while some central and local northern areas show relatively low values. In contrast, the spatial gradients of DMC and DC are clearer and have stronger regional continuity. The study-period mean DMC is about 58–83, with high-value areas mainly concentrated in the Jinan and Jin-southwestern regions, extending in a belt along the central-southern mountains, while the northern and northeastern areas are generally lower. The study-period mean DC is about 223–305, with a wider range of high values, particularly prominent in the western, southern, and some mountainous ridge areas, showing obvious long-term drought accumulation characteristics. As a comprehensive fire risk indicator, the study-period mean FWI is about 4.8–6.3. Its spatial distribution is more consistent with the high-value areas of DMC and DC, generally showing a significant pattern of high in the south and low in the north. Among these, the southwestern and southern boundary areas constitute the main high fire risk core zone of the province, while the northern and northeastern regions maintain relatively low levels. This indicates that, on a average scale, the comprehensive fire risk level in Shanxi Province is not only affected by the rapid drying process of surface fuels but also closely related to the continuous accumulation of medium-deep drought background. At the same time, the relatively fine texture features in the spatial map also suggest that terrain undulations have a significant modulating effect on the distribution of fire risk indices, especially in the mountain-basin transition zone, where fire risk spatial differences are more prominent (Figure 10).
Figure 9
Figure 10
Based on the 2025 study-period mean values of the four indicators (DC, DMC, FFMC, and FWI), coupled zoning of fire weather conditions in Shanxi Province was performed using the K-means clustering method after standardization. Before K-means clustering, the four input variables were standardized with z-score standardization to avoid dominance by variables with larger numerical ranges. We used k = 5 to provide an interpretable five-level fire-weather zoning (low, sub-low, transition, sub-high, and high) consistent with practical risk-management categories. The clustering is used as a descriptive regionalization of the study-period mean fields; it is not presented as a formally optimized or temporally stable classification. The results show that the provincial fire risk background is not uniformly continuous but presents a spatial pattern where significant regional differentiation and gradient transition coexist. Overall, the fire risk level gradually increases from north to south, reflecting consistent changes in deep drought accumulation, duff layer dryness, and fine fuel flammability in space, while also revealing differences in fire risk formation mechanisms in different regions.
From the perspective of spatial distribution, Type I low fire risk areas are mainly concentrated in the northern and northeastern regions. These areas are affected by higher altitudes and relatively humid climate conditions, resulting in lower fuel dryness throughout the year. Type II sub-low fire risk areas are distributed in their periphery, forming a continuous transition zone in the central-north, indicating that fire risk spatial changes have obvious gradual characteristics rather than abrupt processes. Type III transition areas are mainly located in the central and some western and southeastern regions, serving as key units connecting low and high fire risk areas. Various indicators in this region are at medium levels, sensitive to seasonal changes and short-term extreme meteorological processes, reflecting strong spatial heterogeneity. Type IV sub-high fire risk areas are mainly distributed in the central-south and southern peripheral regions, showing enhanced drought accumulation and fuel dryness characteristics, and usually distributed in belts or mosaics around high fire risk areas, forming transitional boundaries for the outward expansion of high-risk areas. Type V comprehensive high fire risk areas are concentrated in the south, especially in the southwestern and southern edge regions, showing a certain scale of contiguous distribution characteristics. This region exhibits high values across all four indicators, indicating its strongest fire risk background, including significant deep drought, low duff layer moisture content, and high flammability of fine fuels, making it the most sensitive core area for provincial fire weather conditions.
From the overall pattern, fire weather conditions in Shanxi Province show a “lower in the north—transitional in the center—higher in the south” spatial structure, superimposed with a certain degree of local patch-like differentiation. The northern region is dominated by low risk, the central region is in a dynamic transition state, while the southern region is dominated by sub-high and high fire risk areas, showing a stronger drought background and fire risk potential. This distribution pattern is in good agreement with regional natural geographical conditions.
It should be pointed out that this coupled zoning result is based on comprehensive multi-indicator information, not a simple mapping of a single FWI. Instead, it is an overall representation of deep fuel drought, duff layer status, fine fuel flammability, and comprehensive fire risk level. Therefore, compared to a single indicator distribution, this method can more comprehensively characterize the spatial structural features of the regional fire weather environment, providing more targeted scientific basis for fire risk zoning management and early warning.
4 Discussion
4.1 Interpretation of temporal and spatial patterns
Through long-term meteorological station observations and multi-source fusion inversion, this study clearly reveals the strong heterogeneity of the forest fire weather index in Shanxi Province across spatiotemporal dimensions. In terms of temporal evolution patterns, the research results indicate that the fire risk potential in Shanxi Province exhibits typical high fire risk characteristics concentrated in late spring and early summer. This aligns with the East Asian monsoon climate and the local vegetation phenological rhythm background of the study area. From the perspective of meteorological forcing, spring (March to May) in Shanxi Province is in the transition period from winter monsoon to summer monsoon, where the Siberian High gradually weakens and the continental thermal low begins to develop. At this time, due to the combined effect of strong cold air remnants and rapid warming, the region experiences a large number of windy days in spring, and the absolute humidity of the air is extremely low. This meteorological combination of high temperature, strong winds, and extreme dryness directly leads to the FFMC, representing surface litter moisture content, frequently exceeding dangerous thresholds in spring.
From an ecological and forest fire behavior perspective, spring is the stage when warm-temperate deciduous broad-leaved forests and mixed coniferous-broad-leaved forests in Shanxi Province are in their unleafed period or early green-up stage. At this time, the forest canopy lacks effective leaf cover, and solar shortwave radiation can directly reach the forest floor, greatly accelerating the physical evaporation of surface litter and dead grass (Maezumi et al., 2024). Simultaneously, the biomass accumulated in the previous year, after experiencing a long dry winter, has reached extremely low moisture content levels, becoming a continuous bed of fine fuels that are highly susceptible to ignition. After entering July and August, concentrated precipitation not only directly eliminates the fire risk potential of FFMC but also causes the lagged-response DC to begin receiving effective moisture replenishment. Coupled with the vegetation entering its vigorous growth period and canopy closure reaching its maximum, the microenvironment within the forest becomes cool and humid, thereby completely ending the high fire risk period for the entire year (Chen et al., 2025).
In terms of spatial distribution patterns, the inversion results show features of high values in the south and low values in the north, with high concentrations at basin edges and on the windward slopes of mountain ranges. This is the result of the combined influence of macro-climatic latitudinal zonation and complex meso- and micro-scale terrain. The main topography of Shanxi Province consists of interweaving mountains on the eastern edge of the Loess Plateau and a series of graben-type fault basins. Due to its lower latitude, the southern region experiences a faster increase in solar altitude angle in spring, receiving more solar radiation than areas like Datong and Shuozhou in the north (Li et al., 2026b). This spatial difference in radiation energy leads to earlier warming in spring and shorter snowmelt periods in the southern region, causing surface fuels to enter a rapid moisture-consuming dry state earlier. Therefore, the high-value accumulation time for FFMC and DMC is much longer than in the north. More noteworthy is the reshaping effect of terrain on local micro-meteorology. Thus, the high fire risk areas appearing in the southern and central-southern mountainous forests and agroforestry intercropping zones on the spatial distribution map are not only an inevitable consequence of climate distribution but also a direct manifestation of complex terrain dynamics and thermodynamic mechanisms in forest fire meteorology.
4.2 Main meteorological controls on FWI in Shanxi Province
A deeper analysis of the driving mechanisms of key meteorological factors on FWI changes is of great significance for enhancing regional forest fire warning capabilities. Based on the statistical correlation analysis and time series characteristics results of this study, wind speed and relative humidity are identified as core controlling factors dominating the occurrence of short-term extreme fire risk events, while air temperature and precipitation primarily determine the long-term background state of regional fire risk. During the short-term fire behavior evolution process, there is an almost instantaneous thermodynamic equilibrium relationship between relative humidity and the moisture content of fine surface fuels (Wepryk et al., 2025).
When low-humidity weather occurs frequently in spring, the equilibrium moisture content (EMC) of surface fuels rapidly decreases, leading to a quick dehydration process driven by a significant vapor pressure deficit. Simultaneously, wind speed, by weakening the near-surface laminar boundary layer structure and enhancing turbulent exchange efficiency, increases the rate of water diffusion, thereby accelerating the fuel drying process. After a fire occurs, wind speed further controls the Initial Spread Index through a nonlinear amplification mechanism. When strong winds and low-humidity conditions act synergistically, even if medium-deep fuels still have a certain moisture content, surface fuels can still rapidly reach a combustible state, thereby supporting the rapid spread of fire. This mechanism can effectively explain the frequent spike-like extreme events observed in the FWI time series during spring, which usually correspond to sudden dry and hot meteorological conditions caused by cold front passages or Foehn processes (Van Wagner, 1987).
Compared to the rapid triggering effect of wind speed and humidity on short-term fire risk, air temperature and precipitation affect the fire’s energy release potential and persistence by regulating the Build-up Index. Precipitation has significant nonlinear characteristics in its impact on FWI at various levels: small amounts of precipitation can rapidly reduce the moisture content of fine fuels, thereby lowering FFMC. However, due to factors such as canopy interception and surface runoff, its replenishment of deep soil moisture is limited, causing DC to remain at a high level. Under climatic conditions of continuous high temperatures and less precipitation, potential evapotranspiration remains high for extended periods, driving continuous dehydration of deep organic matter. When a long-term drought background is superimposed with short-term extreme meteorological conditions, the fire risk system exhibits significant cross-time scale coupling effects, thereby increasing fire intensity and duration. This nonlinear synergistic effect of long-term and short-term meteorological factors is an important physical basis for the occurrence of regional major and extra-large forest fires (Liu et al., 2015).
4.3 Advantages and uncertainties of station-based interpolation
In the process of regional fire risk spatial modeling, this study adopts a multi-source data fusion inversion strategy driven by meteorological station observations as a benchmark. Compared to traditional methods relying on spatial interpolation or direct use of reanalysis data, this approach improves the physical interpretability within the FWI system and enhances its ability to characterize extreme values. Ground meteorological stations, based on high-precision observation instruments, can capture near-surface temperature and humidity changes and extreme wind speed features with high temporal resolution. However, reanalysis data, due to their lower spatial resolution, often exhibit a smoothing effect on extreme meteorological elements under complex terrain conditions, particularly showing a systematic underestimation in characterizing extreme wind speeds and low humidity (Silva et al., 2022).
Furthermore, considering that the FWI system has significant nonlinear response characteristics to input meteorological variables, this bias will be progressively amplified during the physical iteration process, leading to an underestimation of fire risk. By using FWI calculated at stations as supervised information, machine learning models can effectively learn and correct systematic errors in reanalysis data. Simultaneously, introducing multi-source microwave remote sensing soil moisture data as key auxiliary variables helps to enhance the model’s ability to characterize surface moisture states and their dynamic changes, thereby achieving a good balance between spatial continuity and extreme response capability in the inversion results (Emick et al., 2023; Singh et al., 2026).
However, this methodological framework still faces certain uncertainties and limitations. Firstly, there is a significant mismatch between the spatial distribution of meteorological stations and the actual areas where forest fires occur. Existing observation stations are mostly distributed in relatively flat and open areas, and their observation results can hardly represent the true meteorological conditions in complex mountainous environments. Secondly, the spatial heterogeneity of precipitation is difficult to be fully captured by a limited observation network, which may lead to biases in DMC and DC estimates, thereby affecting the model’s judgment of fire risk potential. In addition, under complex terrain conditions, the accuracy of microwave remote sensing soil moisture inversion is affected by factors such as terrain undulations, vegetation structure, and surface roughness, and its errors may be amplified during the multi-source fusion process. The above uncertainties indicate that although this framework has high reliability at the regional scale, further introduction of terrain correction and physical constraint mechanisms is still needed for fine-scale characterization at local scales.
4.4 Limitations and future work
Although this study successfully constructed a high-precision multi-source data fusion FWI spatiotemporal inversion framework, this research system still has certain limitations due to the inherent settings of basic theoretical models and the bottlenecks of current Earth observation technology. The first is the ecological climate universality issue of the Canadian FWI system. The initial physical equations and multi-level parameters of the FWI system, such as drying rate constants, equilibrium moisture content curves, and lag coefficients, were originally derived based on standard flat forest burning experiments in the boreal forest of northern Canada. Directly applying them to Shanxi Province, China, where the geomorphology is fragmented, loess gullies crisscross, and vegetation is dominated by warm-temperate deciduous broad-leaved forests, there must be non-negligible ecological biases in the intrinsic fuel drying and dehydration mechanisms. For example, the water holding capacity, porosity, and aerodynamic roughness of the large amount of broad leaves shed by deciduous broad-leaved forests in autumn and winter are different from the needle litter beds of Canadian coniferous forests. Using the same set of static empirical parameters for long-term iteration may lead to systematic overestimation or underestimation in specific seasons. Passive microwave products represent near-surface soil moisture and do not directly observe the deep organic layer represented by DC. In M3, surface soil moisture is therefore used as an auxiliary covariate summarizing recent land-surface water-balance conditions, not as a direct measurement or physical constraint on DC. The modest DC gain and the absence of a DMC gain indicate that the added information is limited for the longer memory moisture codes.
Secondly, in the current modeling framework, we focused on the fusion of meteorological drivers and surface soil moisture, but did not truly couple dynamic vegetation physical parameters, such as dynamic Leaf Area Index (LAI), live fuel moisture content (LFMC), and fine-grained forest fuel load and fuel classification, into the core algorithm. This makes the model highly accurate in assessing pure meteorological fire risk, but lacks critical fuel bed support when converting to the actual probability of forest fire occurrence. In addition, although the 0.01° × 0.01° spatial resolution meets the needs of regional-scale climatic assessment, it is still somewhat coarse for practical application spatial resolutions (Zhou et al., 2025).
To address the above limitations, future research urgently needs to explore breakthroughs in three dimensions: localization of physical mechanisms, deep coupling of multi-dimensional parameters, and downscaling of spatiotemporal scales. At the theoretical level, there is an urgent need to combine field controlled burning experiments and laboratory thermodynamic analysis in northern China to localize and reconstruct the core empirical parameters of the FWI system, establishing a derived fire risk assessment system truly adapted to the characteristics of warm-temperate deciduous broad-leaved forests and mixed coniferous-broad-leaved forests. In terms of methodological expansion, future models should achieve a leap from purely meteorological fire risk indices to comprehensive forest fire risk probability models. XGBoost treats each station-day as a feature vector and therefore does not explicitly encode spatial adjacency or long-range temporal dependence. This distinguishes it from U-Net-style encoder–decoder models used in radar nowcasting (Ayzel et al., 2020), deep generative and physics-conditioned nowcasting models that retain multiscale precipitation structure (Ravuri et al., 2021; Zhang et al., 2023), graph-based forecast systems that propagate information across multiresolution meshes (Lam et al., 2023), and emerging state-space/Mamba models for spatiotemporal weather forecasting (Xiao et al., 2026). The aim of the present study is not to establish superiority over these neural architectures, but to isolate the incremental contribution of satellite soil-moisture covariates within an interpretable, station-supervised FWI framework. Direct benchmarking would require multi-year records, harmonized predictors and targets, and a common spatiotemporally blocked validation protocol; we therefore identify this as a specific future comparison.
Finally, by leveraging multi-scale geographically weighted regression techniques in spatial statistics or high-resolution reconstruction algorithms based on deep learning, the spatial resolution can be downscaled from 10 km to 1 km or even 100 m, truly achieving precise micro-terrain fire risk characterization at the hilltop and plot level. A further limitation is that the baseline comparison is restricted to ERA5-Land-driven FWI and its machine-learning-corrected variants; no convection-permitting or high-resolution mesoscale simulation (e.g., WRF) was included. ERA5-Land is an approximately 9-km offline land-surface simulation forced by downscaled ERA5 meteorology, and interpolation to the 0.01° working grid cannot recover atmospheric structures absent from that forcing (Muñoz-Sabater et al., 2021). In Shanxi’s complex terrain, unresolved elevation gradients, valley–slope circulations, topographic channeling, and rain-shadow or orographic processes can smooth, attenuate, or spatially displace local extremes in wind, relative humidity, temperature, and precipitation. These errors may propagate nonlinearly through the FWI system: wind and humidity directly affect FFMC and ISI, whereas precipitation errors accumulate in DMC and DC, potentially attenuating or mislocating local mesoscale high-risk events. A WRF-based baseline could represent terrain-driven circulations and precipitation more explicitly, but it would not be error-free because its performance also depends on boundary forcing, terrain resolution, and physical parameterizations (Carvalho et al., 2011; Yang et al., 2025). Therefore, the present results support regional-scale bias correction against station-derived targets, but they do not establish local dynamical fidelity or superiority over high-resolution mesoscale modeling. Future work should compare ERA5-Land, WRF, and the fusion model under a common station-based, spatiotemporally blocked protocol, including event-based metrics for upper-tail wind, humidity, precipitation, and FWI.
5 Conclusion
This study constructs a station-supervised, regional-scale FWI retrieval framework that integrates multi-source remote-sensing predictors and machine learning. The results demonstrate improved agreement with station-derived FWI targets relative to the ERA5-Land-driven baseline, while the absence of a high-resolution dynamical baseline and spatially independent validation limits conclusions about local mesoscale extremes in unsampled mountainous locations. Using quality-controlled ground meteorological observation data as a benchmark, combined with ERA5-Land meteorological variables, terrain information, and fused microwave soil moisture data, the three core moisture codes (FFMC, DMC, and DC) in the FWI system were inverted, and ISI, BUI, and FWI were further derived through physical relationships. The research results indicate that the hybrid modeling strategy proposed in this study has good effectiveness and stability. Compared to direct physical calculation methods based on ERA5, this strategy can improve model accuracy. Among them, the introduction of satellite soil moisture data effectively enhanced the ability to characterize surface water dynamic processes, reduced uncertainties brought by reanalysis data, making the ERA5 + SMC + XGBoost scheme perform optimally overall. From a regional characteristic perspective, fire risk in Shanxi Province exhibits significant seasonal variations and spatial heterogeneity, with high-risk periods mainly concentrated from spring to early summer, and high-risk areas concentrated in the central-south and southwestern mountainous regions, indicating the important role of multi-source moisture information in fire risk assessment in complex terrain.
Statements
Data availability statement
The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.
Author contributions
ZL: Conceptualization, Resources, Writing – review & editing. JizL: Methodology, Supervision, Writing – original draft. JinL: Conceptualization, Data curation, Formal analysis, Funding acquisition, Investigation, Resources, Validation, Visualization, Writing – original draft, Writing – review & editing. HY: Software, Writing – original draft. SW: Validation, Writing – original draft. ZD: Writing – original draft.
Funding
The author(s) declared that financial support was received for this work and/or its publication.
Conflict of interest
SW was employed by State Grid Shanxi Electric Power Company.
The remaining author(s) declared that this work was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.
The authors declare that this study received funding from State Grid Shanxi Electric Power Co., Ltd, Project No. 52053024000W. The funder was involved in data collection and analysis.
Generative AI statement
The author(s) declared that Generative AI was not used in the creation of this manuscript.
Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.
Publisher’s note
All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.
References
1
AtalayH.SunarA. F.DervisogluA. (2024). Investigating FWI moisture codes in relation to satellite-derived soil moisture data across varied resolutions. Fire7:272. doi: 10.3390/fire7080272
2
AyzelG.SchefferT.HeistermannM. (2020). RainNet v1.0: a convolutional neural network for radar-based precipitation nowcasting. Geosci. Model Dev.13, 2631–2644. doi: 10.5194/gmd-13-2631-2020
3
BaoW. L. T. Y.LiuY. X. (2025). Spatial characteristics and influencing factors of traditional villages distribution in the Yellow River Basin. Sustainability17:4834. doi: 10.3390/su17114834
4
CarvalhoA. C.CarvalhoA.MartinsH.MarquesC.RochaA.BorregoC.et al. (2011). Fire weather risk assessment under climate change using a dynamical downscaling approach. Environ. Model. Softw.26, 1123–1133. doi: 10.1016/j.envsoft.2011.03.012
5
ChanS. K.BindlishR.O’NeillP. E.NjokuE.JacksonT.CollianderA.et al. (2016). Assessment of the SMAP passive soil moisture product. IEEE Trans. Geosci. Remote Sens.54, 4994–5007. doi: 10.1109/tgrs.2016.2561938
6
ChenT.GuestrinC. (2016). XGBoost: A Scalable Tree Boosting System. In: Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pp. 785–794.
7
ChenS.JainP.RamseyE.ChenJ.FlanniganM. (2025). Comparative analysis of ensemble and deterministic models for fire weather index (FWI) system forecasting. Weather Forecast.40, 2579–2600. doi: 10.1175/waf-d-25-0069.1
8
De JongM. C.WoosterM. J.KitchenK.ManleyC.GazzardR.McCallF. F. (2016). Calibration and evaluation of the Canadian Forest fire weather index (FWI) system for improved wildland fire danger rating in the United Kingdom. Nat. Hazards Earth Syst. Sci.16, 1217–1237. doi: 10.5194/nhess-16-1217-2016
9
Di GiuseppeF.RémyS.PappenbergerF.WetterhallF. (2018). Using the fire weather index (FWI) to improve the estimation of fire emissions from fire radiative power (FRP) observations. Atmos. Chem. Phys.18, 5359–5370. doi: 10.5194/acp-18-5359-2018
10
EmickE.BabcockC.WhiteG. W.HudakA. T.DomkeG. M.FinleyA. O. (2023). An approach to estimating forest biomass while quantifying estimate uncertainty and correcting bias in machine learning maps. Remote Sens. Environ.295:678. doi: 10.1016/j.rse.2023.113678
11
FaalR.SabooriM.PatroE. R.Ala-AhoP.HaghighiA. T. (2025). A novel spatiotemporal analysis of snow cover pattern over Finland based on ERA5-land reanalysis. J. Hydrol.663:264. doi: 10.1016/j.jhydrol.2025.134264
12
GabbanA.San-Miguel-AyanzJ.ViegasD. X. (2008). A comparative analysis of the use of NOAA-AVHRR NDVI and FWI data for forest fire risk assessment. Int. J. Remote Sens.29, 5677–5687. doi: 10.1080/01431160801958397
13
GaoL.GaoQ.ZhangH.LiX.ChaubellM. J.EbtehajA.et al. (2022). A deep neural network based SMAP soil moisture product. Remote Sens. Environ.277:3059. doi: 10.1016/j.rse.2022.113059
14
GuB. W.KangC. K. Y. (2025). Fire safety and ionic conductivity of ternary electrolyte solvents (EC, DEC, and DMC) in lithium-ion batteries: experimental validation and theoretical modeling. Ionics31, 12717–12726. doi: 10.1007/s11581-025-06762-8
15
IshitobiT.KonoY.MochizukiY. (2023). Modeling framework for batch-dependent dynamics of reaction process by combining first principles and machine learning. Electron. Commun. Jpn.106:428. doi: 10.1002/ecj.12428
16
JiaQ.QuanX.De DiosV. R.YebraM.HeB.LiX.et al. (2026). Enhancing two-week live fuel moisture content forecasts through biophysical modelling and remote sensing data assimilation. Remote Sens. Environ.335:115267. doi: 10.1016/j.rse.2026.115267
17
JiangW. Y.QiaoY. M.ZhengX. X.ZhouJ. H.JiangJ. C.MengQ. X.et al. (2024). Wildfire risk assessment using deep learning in Guangdong Province, China. Int. J. Appl. Earth Obs. Geoinf.128:750. doi: 10.1016/j.jag.2024.103750
18
LamR.Sanchez-GonzalezA.WillsonM.WirnsbergerP.FortunatoM.AletF.et al. (2023). Learning skillful medium-range global weather forecasting. Science382, 1416–1421. doi: 10.1126/science.adi2336,
19
LewisC. H. M.LittleK.GrahamL. J.KettridgeN.IvisonK. (2024). Diurnal fuel moisture content variations of live and dead Calluna vegetation in a temperate peatland. Sci. Rep.14:322. doi: 10.1038/s41598-024-55322-z,
20
LiL. (2025). A comprehensive survey of fire weather index (FWI) systems and IoT applications in peatland fire management. IEEE Access13, 33579–33599. doi: 10.1109/ACCESS.2025.3543570
21
LiY.LiJ.TangJ.TanJ.MaT.LiK. (2026b). Dynamics and driving factors of vegetation carbon sources and sinks in Shanxi Province, China. Int. J. Environ. Sci. Technol.23:7049. doi: 10.1007/s13762-026-07049-9
22
LiuX. P.ZhangJ. Q.TongZ. J. (2015). Modeling the early warning of grassland fire risk based on fuzzy logic in Xilingol, Inner Mongolia. Nat. Hazards75, 2331–2342. doi: 10.1007/s11069-014-1428-5
23
LouL. M.MaW. B.ChengP. L.LiuH.HuangY. (2026). Climatic and fuel drivers of lightning-induced forest fire burned area in the Da Hinggan Ling region, Northeast China. Remote Sens.18:657. doi: 10.3390/rs18040657
24
MaezumiS. Y.FletcherM. S.SaffordH.RobertsP. (2024). Fighting with fire: historical ecology and community-based approaches to fire management, stewardship, and ecosystem resilience. One Earth7, 936–941. doi: 10.1016/j.oneear.2024.05.016
25
MaiJ.ChenY. L.SongQ. H.XuZ. Y.LuD. S.LuoG. P. (2025). Carbon emissions from forest harvest and fire offset approximately half of carbon sequestration of forestation in China during 1986-2020. Agric. For. Meteorol.375:830. doi: 10.1016/j.agrformet.2025.110830
26
MillerG.NealeT.GriffinA. L.DootsonP.BeggC.GardnerA. (2025). Putting predictions in public: professional perspectives on the risks and benefits of changing wildfire warning systems. Environ. Hazards25, 225–246. doi: 10.1080/17477891.2025.2467634
27
Muñoz-SabaterJ.DutraE.Agustí-PanaredaA.AlbergelC.ArduiniG.BalsamoG.et al. (2021). ERA5-land: a state-of-the-art global reanalysis dataset for land applications. Earth Syst. Sci. Data13, 4349–4383. doi: 10.5194/essd-13-4349-2021
28
PanJ. H.WangW. G.LiJ. F. (2016). Building probabilistic models of fire occurrence and fire risk zoning using logistic regression in Shanxi Province, China. Nat. Hazards81, 1879–1899. doi: 10.1007/s11069-016-2160-0
29
RavuriS.LencK.WillsonM.KanginD.LamR.MirowskiP.et al. (2021). Skilful precipitation nowcasting using deep generative models of radar. Nature597, 672–677. doi: 10.1038/s41586-021-03854-z,
30
ShangD. F.ZhangF.YuanD. P.HongL.ZhengH. Z.YangF. H. (2024). Deep learning-based forest fire risk research on monitoring and early warning algorithms. Fire7:151. doi: 10.3390/fire7040151
31
SilvaE. H. D.SilvaF. D. D.Da SilvaR. J.r.PintoD. D. C.CostaR. L.GomesH. B.et al. (2022). Performance assessment of different precipitation databases (gridded analyses and reanalyses) for the new Brazilian agricultural frontier: SEALBA. Water14:1473. doi: 10.3390/w14091473
32
SinghA.NiranjannaikM.GauravK. (2026). Downscaling SMAP soil moisture using a hybrid machine-learning algorithm. Appl. Soft Comput.193:114766. doi: 10.1016/j.asoc.2026.114766
33
Van VlietL.FykeJ.NakonecznyS.MurdockT. Q.JafarpurP. (2024). Developing user-informed fire weather projections for Canada. Clim. Serv.35:505. doi: 10.1016/j.cliser.2024.100505
34
Van WagnerC. E. (1987). Development and Structure of the Canadian Forest Fire Weather Index System. Forestry Technical Report 35. Ottawa, ON: Canadian Forestry Service, 662–15198.
35
Van WilgenB. W. (2025). Fire ecology and management in South Africa: a history of research achievements and future challenges. S. Afr. J. Bot.177, 699–712. doi: 10.1016/j.sajb.2025.01.004
36
WeiT. Y.ZhangT. Y.De BruinK.GlomrodS.ShiQ. H. (2017). Extreme weather impacts on maize yield: the case of Shanxi Province in China. Sustainability9:41. doi: 10.3390/su9010041
37
WeprykO.NiklassonM.NordlindE.DrobyshevI. (2025). Modelling of forest fuel and weather effects on fire behavior in the oak forests of southern Sweden. Scand. J. For. Res.40, 357–370. doi: 10.1080/02827581.2025.2531994
38
XiaoY.JiJ.LeiM.WangM.YuT. (2026). WaveST-mamba: a joint framework of wavelet transform with mamba for stable and fluctuating patterns in spatio-temporal weather forecasting. Eng. Appl. Artif. Intell.166:113684. doi: 10.1016/j.engappai.2025.113684
39
XieH.HanW.BiL. (2023). Assimilating FY3D-MWRI 23.8 GHz observations in the CMA-GFS 4DVAR system based on a pseudo all-sky data assimilation method. Q. J. R. Meteorol. Soc.149, 3014–3043. doi: 10.1002/qj.4544
40
YangK.ChenJ.ZhaoT.LuC.XuX.LuoY.et al. (2025). Effects of fine terrain complexity on cloud and precipitation changes over the Tibetan plateau: a modeling study. NPJ Clim. Atmos. Sci.8:22. doi: 10.1038/s41612-025-00907-2
41
ZhangY.LongM.ChenK.XingL.JinR.JordanM. I.et al. (2023). Skilful nowcasting of extreme precipitation with NowcastNet. Nature619, 526–532. doi: 10.1038/s41586-023-06184-4,
42
ZhouP.LiuL. Z.ZhouM. (2025). Forest fuel load estimation using combined Sentinel-1 and Sentinel-2 data: a case study of the greater Khingan Mountains in China. IEEE Access13, 31121–31130. doi: 10.1109/access.2025.3542128
Summary
Keywords
ERA5-land, fire weather index, microwave remote sensing, multi-model combination, XGBoost
Citation
Lu Z, Liang J, Li J, Yuan H, Wang S and Duan Z (2026) Spatiotemporal variation characteristics and multi-combination modeling assessment of fire weather index in Shanxi Province. Front. For. Glob. Change 9:1892417. doi: 10.3389/ffgc.2026.1892417
Received
27 May 2026
Revised
15 August 2026
Accepted
09 September 2026
Published
30 September 2026
Volume
9 - 2026
Edited by
Aseesh Pandey, Govind Ballabh Pant National Institute of Himalayan Environment and Sustainable Development, India
Updates
Copyright
© 2026 Lu, Liang, Li, Yuan, Wang and Duan.
This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.
*Correspondence: Jinsong Li, ljskssss@sina.com
Disclaimer
All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article or claim that may be made by its manufacturer is not guaranteed or endorsed by the publisher.