A Spatiotemporal Wildfire Risk Prediction Framework Integrating Density-Based Clustering and GTWR-RFR

Abstract

Accurate wildfire prediction and identification of key environmental drivers are critical for effective wildfire management. We propose a spatiotemporally adaptive framework integrating ST-DBSCAN clustering with GTWR-RFR. In this hybrid model, Random Forest captures local nonlinear relationships, while GTWR assigns adaptive spatiotemporal weights to refine predictions. Using historical wildfire records from Hunan Province, China, we first derived wildfire occurrence probabilities via ST-DBSCAN, avoiding the need for artificial non-fire samples. We then benchmarked GTWR-RFR against seven models, finding that our approach achieved the highest accuracy (R² = 0.969; RMSE = 0.1743). The framework effectively captures spatiotemporal heterogeneity and quantifies dynamic impacts of environmental drivers. Key contributing drivers include DEM, GDP, population density, and distance to roads and water bodies. Risk maps reveal that central and southern Hunan are at high risk during winter and early spring. Our approach enhances both predictive performance and interpretability, offering a replicable methodology for data-driven wildfire risk assessment.

1. Introduction

Wildfires increasingly threaten forest ecosystems and human settlements worldwide, with rising frequency and severity under climate change [1,2]. Hunan Province, China, is especially vulnerable due to its dense vegetation, complex terrain, and frequent human activities, leading to substantial ecological and economic losses [3]. Accurate wildfire prediction remains challenging because of the complex, nonlinear interactions among climatic, topographic, vegetative, and anthropogenic factors [4].

Traditional “black-box” machine learning models (e.g., RF, SVM, MaxEnt) achieve high accuracy but assume global stationarity and overlooking local spatiotemporal dynamics [5,6,7]. Spatial regression techniques like Geographically Weighted Regression (GWR) and Geographically and Temporally Weighted Regression (GTWR) address non-stationarity by allowing coefficients to vary over space and time [8,9,10], but their linear nature constrains its capacity to model complex nonlinear interactions inherent in wildfire dynamics and may lead to multicollinearity and overfitting [11,12]. Recent studies have proposed hybrid frameworks that integrate machine learning algorithms with spatial regression techniques to flexibly adjust spatiotemporal weights and better capture spatiotemporal non-stationarity [13]. Among them, the GWR-RF model incorporates RF into the GWR structure, enhancing both predictive performance and interpretability [14] in domains such as traffic risk assessment [15], crop yield prediction [16], and disaster early warning [17]. However, its application to wildfire risk remains scarce.

In addition to model formulation, the construction of the dependent variable remains an underexplored aspect of wildfire modeling. Conventional binary classification relies on artificially generated non-fire samples, which can distort spatial dependencies. In contrast, ST-DBSCAN, a spatiotemporal extension of DBSCAN, identifies meaningful wildfire clusters by jointly considering spatial and temporal proximity [18,19], thereby preserving the intrinsic structure of wildfire occurrences [20,21].

To address these limitations, this study proposes a novel integrated wildfire risk prediction framework that combines ST-DBSCAN clustering with a Geographically and Temporally Weighted Random Forest Regression (GTWR-RFR). ST-DBSCAN identifies realistic wildfire clusters without artificial controls, preserving spatiotemporal dependency, while GTWR-RFR combines Random Forest’s nonlinear learning with GTWR’s spatiotemporal weighting. The objectives are to: (1) estimate wildfire occurrence probability using ST-DBSCAN to avoid sampling bias; (2) compare GTWR-RFR with conventional machine learning and spatial models; and (3) reveal the spatiotemporal effects of environmental and socioeconomic drivers for localized wildfire risk assessment. By leveraging Random Forest’s nonlinear modeling and GTWR’s adaptive spatiotemporal weighting, the proposed framework aims to deliver localized, interpretable, and high-fidelity wildfire risk assessments.

2. Study Area and Data Preparation

2.1. Study Area

Hunan Province (24°38′–30°08′ N, 108°47′–114°15′ E) is located in south–central China and spans approximately 211,829 km² (Figure 1). The terrain gradually transitions from low-lying plains in the north, such as the Dongting Lake area, to hilly and mountainous regions in the central and southern parts of the province. Hunan experiences a humid subtropical monsoon climate, characterized by a pronounced dry season from December to February and a wet season from May to August. Forests cover nearly 60% of the province’s land area, with denser vegetation concentrated in the mountainous western and southern zones (Figure 2b). The complex interplay among diverse topography, microclimates, vegetation types, and anthropogenic activities leads to significant spatial heterogeneity in wildfire risk. These characteristics make Hunan a representative region for investigating spatiotemporal wildfire risk under complex environmental and anthropogenic influences.

2.2. Data Collection and Preprocessing

Wildfire records from 2000 to 2024 were obtained from NASA’s Fire Information for Resource Management System (FIRMS), with a spatial resolution of 1 km². To ensure high data reliability, only wildfire events with a confidence level exceeding 70% were retained, and those occurring in non-forested areas were excluded. This filtering process yielded 18,784 high-confidence wildfire records (Figure 2a).

Based on prior studies, 18 explanatory variables were selected to represent meteorological, topographic, vegetation, infrastructure, and socioeconomic drivers of wildfire occurrence (Figure 3 and

Table 1. Data sources and spatial resolution of explanatory variables.

Factors	Variable	Source	Classes	Spatial Resolution/ Units
Wildfire point	Historical Wildfires	NASA Fire Information Resource Management System (FIRMS) https://firms.modaps.eosdis.nasa.gov/ (accessed on 21 October 2025)	Continuous
Meteorological factors	Temperature Pressure Humidity Wind speed Precipitation Sunshine	High-resolution gridded datasets https://crudata.uea.ac.uk/cru/data/hrg/ (accessed on 21 October 2025) China Geospatial Data Cloud http://data.cma.cn/ (accessed on 21 October 2025)	Continuous	0.1 °C 0.1 hPa 1% 0.1 m/s 0.1 mm 0.1 h
Terrain factors	DEM Slope Aspect	Geospatial Data Cloud https://www.gscloud.cn/ (accessed on 21 October 2025)	Continuous	1000 m
Vegetation factors	Vegetation type	Geospatial Information Authority of Japan (GSI) https://globalmaps.github.io/glcnmo.html (accessed on 21 October 2025)	Categorical	1000 m
	NDVI	MODIS Vegetation Index Products https://modis.gsfc.nasa.gov/data/dataprod/mod13.php (accessed on 21 October 2025)	Categorical
	Canopy closure timber volume	National Smart Forest Resources Management Platform https://www.stgz.org.cn/ldbggzpt/ (accessed on 21 October 2025)	Continuous
Anthropogenic factors	Road Water Railway	Openstreetmap https://www.openstreetmap.org/ (accessed on 21 October 2025)	Continuous	1000 m
Socioeconomic factors	GDP	Resource and Environmental Science and Data Center, https://www.resdc.cn/ (accessed on 21 October 2025)	Continuous	1000 m
	Population	WorldPop https://hub.worldpop.org/ (accessed on 21 October 2025)	Continuous

). These variables were chosen based on their documented relevance to wildfire occurrence and availability across the study region. Data were obtained from publicly available remote sensing and national statistical sources. All variables were resampled to a uniform 1 km grid to ensure spatial consistency. Missing values were addressed using inverse distance weighted (IDW) interpolation and temporal averaging, ensuring a complete and harmonized dataset for modeling.

3. Methodology

3.1. Research Process

The research framework integrates data pre-processing, variable selection, spatiotemporal clustering, and model training to assess wildfire risk in Hunan Province. The complete research workflow is illustrated in Figure 4.

Following data filtering, ST-DBSCAN was applied to identify spatiotemporal clusters of wildfire events and to derive wildfire occurrence probabilities. Environmental variables were subsequently extracted to each wildfire event to construct the model input dataset. All variables were standardized to improve computational stability and model convergence. Spatial autocorrelation analysis was performed to examine underlying spatial patterns, and multicollinearity among explanatory variables was addressed using the Variance Inflation Factor (VIF) analysis.

Eight models were compared in this study: KNN, MLP, SVR, AdaBoost, RFR, GWR, GTWR and GTWR-RFR. To ensure optimal performance and comparability, each model was tuned using grid search and cross-validation. KNN models were optimized by varying the number of neighbors (K); RF and AdaBoost by adjusting tree depth, estimators, and learning rate; MLP by tuning hidden layers, learning rate, and activation function; and SVR by selecting kernel types and adjusting penalty (C) and kernel coefficient (γ). For spatial models, the GWR bandwidth was selected via cross-validation. GTWR, GTWR-RFR required tuning both spatial bandwidth and the temporal scale parameter (μ), with sensitivity analysis conducted over μ values from 0.1 to 0.9 to evaluate model stability and fit. Model performance was evaluated using RMSE and R². GTWR-RFR achieved the lowest RMSE and highest R², and was thus selected as the optimal model. Finally, IDW interpolation was used to visualize the spatial distribution of predicted wildfire probabilities and support subsequent spatial pattern analysis.

3.2. Estimating Wildfire Probability via Spatiotemporal Clustering

To estimate wildfire occurrence probability, this study employed the ST-DBSCAN algorithm, which extends traditional DBSCAN by incorporating both spatial and temporal proximity, without requiring prior specification of the number of clusters [18,22]. For each wildfire event $e_i$, its spatiotemporal neighborhood is defined by events within a spatial radius (${ϵ}_s$) and a temporal window (${ϵ}_t$). If the number of events in this neighborhood meets or exceeds a minimum density threshold (MinPts), $e_i$ is considered a core point of a cluster. The wildfire probability $P_i$ at $e_i$ is then calculated by normalizing its neighborhood density against the maximum density observed across all events:

$$P_i={\displaystyle \frac{\left|N_{({ϵ}_s,{ϵ}_t)}\right(e_i\left)\right|}{{max}_{\mathrm{e}\in D}\left(\left|N_{({ϵ}_s,{ϵ}_t)}\right(e\left)\right|\right)}}$$

(1)

This normalization yields a relative spatiotemporal density scaled between 0 and 1, representing the likelihood of wildfire occurrence. Clustering validity was assessed using the Silhouette Coefficient [23].

3.3. Geographically and Temporally Weighted Random Forest Regression and Evaluation Metrics

The GTWR-RFR model synergistically integrates the adaptive spatiotemporal weighting of GTWR with the robust nonlinear learning capability of RF [14]. While GTWR effectively captures spatiotemporal heterogeneity by allowing coefficients to vary across space and time [24], its linear formulation is sensitive to outliers and incapable of modeling complex nonlinear relationships [11,25]. Conversely, standard Random Forest regression (RFR), though robust and non-parametric, inherently lacks mechanisms to account for spatial and temporal dependencies. GTWR-RFR addresses these limitations by training localized RFR models at each spatial location, weighted according to spatiotemporal proximity. The GTWR model at a spatiotemporal point $\left(u_i,v_i,t_i\right)$ is defined as follows:

$$Y_i={\beta }_0\left(u_i,v_i,t_i\right)+\sum _{k=1}^p{\beta }_k\left(u_i,v_i,t_i\right)X_{ik}+{\epsilon }_i,i=\mathrm{1,2}\dots ,n$$

(2)

where ${\mathsf{\beta }}_k\left(u_i,v_i,t_i\right)$ are the location- and time-specific coefficients, and ${\epsilon }_i~\mathrm{N}(0,{\mathsf{\sigma }}^2)$ are random error term. The coefficients are estimated using weighted least squares:

$$\widehat{\beta }={\left(X^T{W}_{i}X\right)}^{-1}{X}^T{W}_{i}Y$$

(3)

here $W_i=diag(w_{i0},w_{i1},w_{i2},\dots ,w_{in})$ is the spatotemporal weight matrix defined by a Gaussian kernel. The weight $w_{ij}$ between observations i and j is determined by their spatiotemporal distance, calculated as follows:

$$d_{ij}^{ST}=\sqrt{\lambda \left[{\left(u_i-u_j\right)}^2+{\left(v_i-v_j\right)}^2\right]+\mu {(t_i-t_j)}^2}$$

(4)

The corresponding spatiotemporal Gaussian weight function [26] is as follows:

$$w_{ij}^{ST}=\mathit{exp}\left(-{\displaystyle \frac{({d_{ij}^{ST})}^2}{h^2}}\right)=exp\left\{-{\displaystyle \frac{\lambda \left[{\left(u_i-u_j\right)}^2+{\left(v_i-v_j\right)}^2\right]+\mu {\left(t_i-t_j\right)}^2}{h^2}}\right\}$$

(5)

where h is the bandwidth, and λ, μ are scale parameters balancing spatial and temporal influences. The optimal bandwidth $h$ is selected by minimizing the corrected Akaike Information Criterion (AICc).

The spatiotemporal weights from GTWR are used to train local RFR models. The final prediction ${\widehat{y}}_i$ at location iis the average output of all $M$ decision trees in the local forest, incorporating the GTWR-derived coefficients as features:

$${\widehat{y}}_i={\displaystyle \frac{1}{M}}\sum _{m=1}^M{f}_m({\beta }_0^{*},{\beta }_1^{*},\dots ,{\beta }_p^{*},X_i)$$

(6)

where $f_m$() is the prediction function of the m-th tree, ${\beta }_p^{*}$ are the spatiotemporal coefficients from GTWR, and $X_i$ is the vector of explanatory variables at i.

Model performance was evaluated using four standard metrics, including RMSE, MAE, R², and AICc. Lower RMSE and MAE indicate better prediction accuracy, while higher R² reflects stronger explanatory power.

4. Results

4.1. Spatial Characteristics and Feature Selection

Wildfire events in Hunan Province exhibit distinct temporal and spatial heterogeneity. The annual frequency of wildfires rose steadily from 2000 to 2008, followed by a declining trend (Figure 5a). Seasonally, two major wildfire seasons are observed annually, peaking between January–April and October–December (Figure 5b). February records the highest frequency, driven by dry conditions, intermittent winds, and human activities such as traditional festivals. Conversely, wildfire activity is minimal from May to September, when plum rains and summer monsoons elevate vegetation moisture.

Using a 5 km spatial radius and a 3-day temporal window, ST-DBSCAN identified 3347 high-density wildfire clusters, while 29.03% of events were classified as noise, indicating isolated low-risk occurrences. Wildfire occurrence probabilities were then computed for each event based on its cluster (Figure 5c). The log-transformed wildfire occurrence probability exhibits a distribution close to normality (Figure 6a).

Global Moran’s I (0.87; Z = 306.8) confirms strong positive spatial autocorrelation (Figure 6b). Local Indicators of Spatial Association (LISA) further reveal a clear north–south gradient in wildfire occurrence: hotspots concentrate in southern and south-eastern Hunan, whereas coldspots dominate the north and central regions (Figure 6c,d). This significant positive spatial autocorrelation underscores the need for models that incorporate spatial heterogeneity.

To ensure reliable model inputs, multicollinearity among explanatory variables was assessed via variance inflation factor (VIF) analysis [27]. Factors with VIF values between 2.04 and 5.99 were retained (

Table 2. VIF analysis for variable selection.

Model Variable	Variable Code	VIF	TOL
Canopy Closure	CC	0.171556	5.829003
Daily mean temperature	DMT	0.171699	5.824144
DEM	DEM	0.173677	5.75782
Daily mean wind speed	DMWS	0.203361	4.917362
Slope	Slope	0.211842	4.72051
Aspect	Aspect	0.25988	3.847925
Distance from water	NDW	0.269735	3.707344
24 h precipitation	HP	0.287977	3.472499
Distance from road	NDR	0.314261	3.18207
Population density	PD	0.382188	2.616512
Standing timber volume	AVST	0.390855	2.558495
GDP	GDP	0.420837	2.376215

), satisfying VIF < 10 and tolerance > 0.1 criteria [28,29]. The final set of 13 variables captures the principal climatic, topographic, vegetative, and anthropogenic determinants of wildfire risk.

4.2. Model Performance Evaluation

We evaluated eight models, including five machine learning algorithms (KNN, SVR, MLP, AdaBoost, RFR) and three spatial regression techniques (GWR, GTWR, GTWR-RFR), for wildfire probability prediction.

Table 3. Model performance metrics (R², RMSE, MAE, etc.).

	KNN	MLP	SVR	AdaBoost	RFR	GWR	GTWR	GTWR-RFR
R2	0.376	0.532	0.478	0.657	0.802	0.848	0.885	0.969
Explained Variance Score	0.271	0.3697	0.3396	0.5209	0.5709	0.8481	0.8854	0.9696
Mean Absolute Error	0.6647	0.6084	0.619	0.5439	0.5035	0.2663	0.2255	0.1256
Mean Squared Error	0.7457	0.6408	0.6705	0.49	0.4368	0.1521	0.1147	0.0304
Root Mean Squared Error	0.8635	0.8005	0.8188	0.71	0.6609	0.3899	0.3387	0.1743
Explained sum of squares	1061.9	1671.5	1249.6	1765.9	1813.7	14,640.2	15,486.2	15,661.7
Total sum of squares	3813.5	3813.5	3813.5	3813.5	3813.8	18,783	18,783	18,783.0
Residual sum of squares	2801.5	20,606.9	19,022.1	20,961.9	21,141.3	2856.1	2154.8	570.8

summarized the performance metrics of all models. The pronounced performance disparities among these models can be primarily attributed to their inherent capacities for handling spatial heterogeneity and nonlinear relationships. Predicted versus observed probabilities are shown in scatterplots (Figure 7). Among the machine learning methods, RFR achieved the best performance (R² = 0.802; RMSE = 0.6609), whereas KNN performed the poorest (R² = 0.376; RMSE = 0.8635). The superior performance of RFR is attributed to its ensemble structure and intrinsic capability to model complex nonlinear interactions without a priori assumptions. In contrast, the poor performance of KNN likely stems from the curse of dimensionality, which impairs its ability to discern meaningful patterns in high-dimensional feature space [30].

To account for spatial heterogeneity, GWR was fitted using a fixed Gaussian kernel with a bandwidth of 7 639 m, yielding R² = 0.848 and RMSE = 0.3899 (Figure 7f). Incorporating temporal heterogeneity, the GTWR model applied the same spatial bandwidth and a temporal decay parameter (μ = 0.2), improving model performance to R² = 0.885 and RMSE = 0.3387 (Figure 7g). The improvement from GWR to GTWR highlights the importance of addressing temporal non-stationarity alongside spatial variation. GTWR’s adaptive spatiotemporal weighting dynamically adjusts local fits based on proximity in both space and time, effectively representing non-stationary relationships in wildfire drivers [24]. The hybrid GTWR-RFR model was then developed by integrating RFR into the GTWR framework. By combining RF’s nonlinear learning and feature selection capabilities with GTWR’s adaptive spatiotemporal weighting, GTWR-RFR achieved superior predictive accuracy, with R² = 0.969 and RMSE = 0.1743 (Figure 7h). This integration enables the model to capture complex nonlinear wildfire patterns more effectively [31,32]. By leveraging spatially and temporally localized learning, GTWR-RFR effectively resolves the limitations associated with global-stationarity assumptions in conventional machine learning models while overcoming the linear constraints of standard spatial regression approaches.

We applied Global Moran’s I to analyze the spatial autocorrelation of prediction residuals across all models to assess whether the models effectively captured spatial heterogeneity [16,33].

Table 4. Global Moran’s I of residuals for all models.

Model	Moran’s I	Z-Score	p-Value
AdaBoost	0.6050	213.2841	0
SVR	0.6370	224.5765	0
MLP	0.5692	200.6862	0
KNR	0.6618	233.3181	0
RFR	0.5781	203.8316	0
GWR	0.3459	121.9581	0
GTWR	0.2959	104.3545	0
GTWR-RFR	0.2462	86.8203	0

summarizes the Moran’s I results. Residual clustering is evident in non-spatial models, with Moran’s I exceeding 0.5. In contrast, GTWR-RFR achieved the lowest value (0.2462), indicating minimal residual autocorrelation and superior spatial generalization. The analysis of residual spatial autocorrelation further quantifies the uncertainty associated with ignoring spatial effects, which signifies unexplained spatial structure and potential model misspecification.

4.3. Spatial Patterns of Coefficient Estimates

Statistical analysis on the GTWR-RFR coefficient estimates is summarized in

Table 5. Descriptive statistics of GTWR-RFR coefficient estimates.

Model Variable	Code	Mean	STD	Min	Median	Max	Moran’s I	z-Score	p-Value
Daily mean temperature	DMT	0.044	0.028	0.003	0.038	0.529	0.8002	282.2	0
24 h precipitation	HP	0.046	0.031	0.003	0.04	0.544	0.8471	298.88	0
Daily mean wind speed	DMWS	0.056	0.037	0.003	0.049	0.682	0.8127	286.6	0
Slope	Slope	0.054	0.032	0.003	0.048	0.365	0.8525	300.6	0
Altitude	DEM	0.128	0.097	0.004	0.097	0.764	0.8454	298.1	0
Aspect	Aspect	0.047	0.027	0.002	0.043	0.484	0.8598	303.3	0
Canopy Closure	CC	0.026	0.021	0.001	0.022	0.51	0.8087	285.6	0
Standing timber volume	AVST	0.022	0.021	0	0.017	0.445	0.8535	301.3	0
GDP	GDP	0.152	0.117	0	0.117	0.795	0.8382	295.5	0
Population density	PD	0.129	0.103	0.007	0.095	0.821	0.8685	306.2	0
Distance from road	NDR	0.11	0.105	0.005	0.078	0.85	0.8961	316.0	0
Distance from water	NDW	0.14	0.111	0.006	0.102	0.784	0.8998	317.2	0

, and most coefficient means are near zero, indicating that local rather than global factors predominantly govern wildfire risk. The standard deviation (STD) of each coefficient quantifies spatial variability: lower STD values (e.g., CC, STD = 0.021) denote consistent influences across Hunan, whereas higher STD values (e.g., GDP, STD = 0.117) reflect substantial local variation in driver effects.

Figure 8 and Figure 9 present IDW-interpolated maps of local variable importance and corresponding t-value distributions, respectively, highlighting spatial heterogeneity in driver impacts on wildfire probability. Among all factors, DEM, GDP, population density, distance to roads, and distance to water bodies emerge as especially influential. DEM governs microclimatic conditions and vegetation distribution, thereby affecting fuel availability and dryness [34,35,36]. Regional GDP reflects economic disparities that shape infrastructure and wildfire prevention investments, with spatial heterogeneity in GDP driving variation in ignition sources and suppression capacity [37]. High population density contributes to forest fragmentation and increased edge effects, thereby elevating wildfire susceptibility [38]. Greater distance from roads correlates with higher risk due to delayed fire-fighting response [39], while increased distance from water bodies corresponds to elevated risk, as water features serve as natural firebreaks [40]. Collectively, these drivers interact through ecological and socioeconomic pathways to shape the complex spatial patterns of wildfire occurrence. These spatially heterogeneous coefficients offer valuable insights for tailoring region-specific wildfire prevention strategies.

4.4. Risk Classification and Spatial Distribution

Based on GTWR-RFR results, wildfire risk in Hunan Province was categorized into four levels (

Table 6. Pixel count and percentage by wildfire risk category.

Risk Classification	Fire Probability	Pixels Count	Percent (%)	Recommended Measures
Non-fire risk	<0.2	3,050,513	45.01%	negligible
Low-fire risk	0.2~0.4	2,293,834	33.85%	routine prevention and monitoring
Moderate-fire risk	0.4~0.6	884,769	13.06%	proactive assessments
High-fire risk	0.6>	547,788	8.08%	Intensive monitoring and targeted prevention strategies

). The spatial distribution reveals a southeast-to-northwest decreasing gradient (Figure 10), aligning with prior studies [41,42]. This pattern reflects the combined influence of topography, climate, vegetation, and human activities on wildfire occurrence, especially in rugged, densely forested terrain with frequent anthropogenic disturbances [43,44]. In the northern and western regions, such as the Dongting Lake basin and the western highlands, wildfire risk remains low due to sparse vegetation, high precipitation, limited fuel, and reduced human disturbance associated with elevation and remoteness. In contrast, high-risk zones are concentrated in the south-eastern, central, and southern regions. The combination of dense vegetation, abundant fuels, high population density, and well-developed road accessibility in these areas significantly elevates the risk of wildfires [45]. Moreover, complex terrain and monsoonal winds in the south-eastern mountains further exacerbate wildfire susceptibility [46,47].

We employed the GTWR-RFR model to generate monthly wildfire risk predictions for Hunan Province, China, revealing distinct and pronounced seasonal variability patterns. The monthly risk maps (Figure 11) clearly illustrate this regularity. High-risk periods occur from January to April and October to December. Winter–spring months (January–March) feature low precipitation and fuel desiccation, particularly in southern and eastern Hunan, with April peaking due to Qingming Festival burning activities. By contrast, May–September shows low risk despite high temperatures, as monsoonal rains maintain high fuel moisture. Risk rebounds in October–December as rainfall wanes and agricultural burning resumes. These findings highlight the necessity of seasonally targeted prevention measures, especially limiting human ignitions, raising public awareness, enhancing wildfire monitoring and optimizing resource deployment, during peak wildfire seasons.

5. Discussion

5.1. Advantages of Probability-Based Wildfire Modeling via ST-DBSCAN

Traditional wildfire risk models typically adopt a binary fire–nonfire classification framework, which often requires artificially generated non-fire samples. These samples tend to share similar environmental characteristics with fire-prone areas, such as climate, topography, vegetation, and human activity, thereby introducing bias [48]. Moreover, binary classification fails to capture the spatially coherent nature of wildfire risk [49,50].

To overcome these limitations, this study employs a probability-based modeling framework. Wildfire probability is estimated from the density of historical wildfire clusters, which are effectively identified by ST-DBSCAN based on spatial and temporal proximity, allowing the distinction between fire-prone areas and isolated occurrences [18,51]. The probability-based framework offers several advantages: (1) Reduces sampling bias by avoiding randomly generated non-fire samples; (2) Preserves spatiotemporal patterns of wildfire occurrence, enabling more interpretable modeling; and (3) Improves data quality by filtering noise and outliers through clustering. This probability-based approach lays a robust foundation for downstream spatial modeling, facilitating improved generalization and interpretability.

5.2. Strengths of the GTWR-RFR Hybrid Regression

Because wildfire probability exhibits strong spatial autocorrelation, adaptive models with spatially varying coefficients are required [52,53]. GTWR-RFR integrates the nonlinear learning capability of RFR with the spatiotemporal weighting mechanism of GTWR, thereby addressing the limitations of both linear spatial models (e.g., sensitivity to outliers) and non-spatial models (e.g., neglect of spatial variation) [54]. By calibrating local parameters based on spatiotemporal proximity, this approach flexibly adjusts weights [10,55], thereby more effectively capturing region-specific variations in key wildfire drivers (e.g., DEM, GDP, population density, distance to roads and water bodies) while retaining model interpretability [24].

GTWR-RFR enhances wildfire risk prediction by capturing both spatial heterogeneity and nonlinearity. Our results demonstrate that it outperforms both non-spatial machine learning algorithms and traditional spatial regression models, achieving superior model fit, the lowest error, and the smallest residual, indicating the highest predictive accuracy and minimized spatial bias. The strong agreement between predicted and observed values validates the model’s robustness. These strengths establish GTWR-RFR as a robust and interpretable framework for capturing dynamic wildfire patterns in heterogeneous environments.

5.3. Limitations and Future Directions

Although the GTWR-RFR framework demonstrates significant advantages in modeling complex spatiotemporal wildfire risk, several limitations remain. First, the model involves high computational complexity. GTWR requires constructing a local weight matrix and estimating regression coefficients for each spatiotemporal sample, while the RF relies on extensive ensemble operations. Second, the model’s performance is sensitive to data quality. Spatially or temporally uneven data distribution may induce local overfitting in the regional RF models, thereby compromising predictive reliability. Finally, the framework’s performance is sensitive to parameterization. Suboptimal spatiotemporal thresholds in ST-DBSCAN or inappropriate RF hyperparameters may induce instability and degrade accuracy [23,56]. Model generalization may be limited by regional or temporal heterogeneity, reducing transferability across study areas.

Future work should prioritize enhancing the model’s computational efficiency, generalization capability, and predictive accuracy. Developing parallel and distributed computing strategies to accelerate processing; exploring alternative regressors such as local Support Vector Machines (SVMs) or neural networks to further increase flexibility and predictive performance; and incorporating high-resolution and spatiotemporally consistent variables (e.g., fuel moisture, fuel load, and fine-scale socioeconomic indicators) to refine model realism.

6. Conclusions

This study developed an integrated spatiotemporal framework combining ST-DBSCAN clustering with GTWR-RFR for wildfire risk prediction. The model significantly improves predictive accuracy by leveraging RF’s nonlinear modeling capability and GTWR’s adaptive spatiotemporal weighting, while also eliminating the bias in artificially generated non-fire samples through density-based probability estimation. Applied to Hunan Province, the framework reveals pronounced spatiotemporal heterogeneity in key wildfire risk drivers. The resulting risk maps and the spatiotemporal heterogeneity of key drivers provide actionable insights for optimizing resource allocation and targeted prevention strategies. Future research should focus on enhancing computational efficiency through parallel and distributed computing, improving predictive flexibility via alternative regressors such as local SVM or neural networks, and extending the framework to diverse geographic regions and climate scenarios to maximize its applicability in wildfire management and early warning systems.