High-Resolution Burned-Area Mapping and Vegetation Resilience in Heterogeneous Landscapes Using Sentinel-2 and Explainable Machine Learning

Abstract

Accurate wildfire impact assessment and understanding post-disturbance recovery are essential for land management in fire-prone regions. This study develops a Sentinel-2–based burned-area extraction framework and integrates NDVI time-series analysis with explainable machine learning to quantify vegetation resilience across five fire-affected regions in China. The burned-area map achieves an overall accuracy of 99.8%, substantially outperforming MODIS products (77.9% and 92.7%) by better detecting fragmented patches in complex terrain. NDVI trajectories reveal three resilience pathways: compensatory recovery, stable recovery without compensation, and persistent degradation. Recovery times ranged from approximately 2 months to over 6 months, with some high-altitude areas showing no effective recovery. An XGBoost–SHAP model explains spatial recovery variability (R² = 0.50–0.88) and identifies a consistent shift from early climate control to later topographic regulation. Landscape heterogeneity promotes recolonization only within intermediate thresholds, temperature exhibits optimal windows, and precipitation shows diminishing returns. Topography acts primarily by redistributing hydrothermal conditions rather than as an independent driver. The results demonstrate strong spatial variability in ecosystem stability and highlight nonlinear interactions among climate, terrain, and landscape structure as key determinants of resilience. The proposed framework improves burned-area monitoring and supports targeted ecological restoration and adaptive land-use planning in heterogeneous landscapes.

1. Introduction

In recent years, driven by the cumulative impacts of human activities and intensified global climate change, extreme high-temperature and drought events have occurred with a markedly higher frequency [1], leading to a notable rise in global wildfire occurrences, fire intensity, and affected areas [2,3,4]. Severe wildfires have caused massive casualties and economic losses worldwide, as evidenced by the 2016 Fort McMurray wildfire in Canada [5], the 2019–2020 Black Summer bushfires in Australia [6], and the 2023 Maui wildfires in the United States [7], highlighting wildfires as a critical climate risk with high suddenness and destructiveness. Therefore, understanding and managing wildfires has become a pivotal issue for global ecological management and public safety.

Wildfires are among the most destructive natural disasters worldwide. They not only directly destroy vegetation structures, alter community composition and impair vegetation recovery capacity [8], but also cause degradation of soil physical and chemical properties, reduction in water-holding capacity and damage to microhabitats [9,10,11,12,13]. In addition, massive carbon emissions from high-intensity fires convert key vegetation-covered areas from carbon sinks to carbon sources, thereby creating a positive feedback loop on global warming [14,15,16,17]. Wildfires also significantly alter regional landscape patterns by accelerating community succession, inducing land cover conversion and ecosystem function degradation. Their superposition with stressors such as droughts, plant diseases and insect pests further amplifies ecological risks [18,19,20,21,22]. Thus, timely identification of fire spots is a core step in wildfire suppression and risk control [23], while post-fire vegetation recovery monitoring is critical to ecosystem stability, habitat restoration and disaster prevention planning [24,25]. The speed, pathways and spatial heterogeneity of post-fire recovery not only reflect ecosystem vulnerability, but also directly determine soil erosion risks, biodiversity restoration potential and regional ecological functions [26]. Hence, developing a rapid, accurate and scalable system for wildfire identification and recovery monitoring is of fundamental importance for ecological conservation and policy formulation [27,28,29].

Remote sensing provides large-scale, continuous, and multi-scale observation data, serving as the core technical means for wildfire monitoring [30,31]. However, traditional satellite data have notable limitations in spatio-temporal resolution, spectral sensitivity, and adaptability to complex surface conditions, restricting the fine characterization of fire dynamics and post-fire ecological recovery. For example, the Landsat series, the longest-running medium-resolution data archive, can capture spatial heterogeneity within fire zones [32,33,34], but its 16-day revisit interval reduces the probability of acquiring cloud-free observations at critical time points [35]. In contrast, MODIS and AVHRR sensors feature ultra-high temporal resolution for large-scale vegetation monitoring and fire severity assessment [36,37,38]. Yet their low spatial resolution generates extensive mixed pixels in small fire patch monitoring, causing misjudgment and hindering fine characterization of within-fire recovery patterns [32,34]. Early datasets such as AVHRR also suffer from systematic noise from orbital drift and sensor aging [37].

In response to the above limitations of multi-source data, Sentinel-2 satellite imagery offers significant advantages, with 10–20 m high spatial resolution reducing cloud interference and mixed pixel effects, a short revisit cycle improving image acquisition probability during critical fire periods, and red-edge bands enabling sensitive capture of vegetation health changes [39]. Relevant research shows its 20 m resolution accurately detects small fire patches below 100 hectares. These patches are often missed by coarse-resolution sensors. This greatly improves the precision of burned area mapping [40]. In addition, time-series Sentinel-2 NDVI sequences enable fine detection of post-fire vegetation changes. Combined with machine learning methods such as linear regression and random forest classification, they achieve reliable and refined monitoring of post-fire recovery [41].

Meanwhile, despite the progress in burned area indices and time series analysis methods, critical research gaps remain. For burned area extraction, the widely used Normalized Difference Vegetation Index (NDVI) is prone to saturation in high-biomass areas and high sensitivity to soil background [42]. To address this, existing studies have proposed the Normalized Burn Ratio (NBR) based on reflectance differences between near-infrared (NIR) and shortwave infrared (SWIR) bands [43]. Nevertheless, high fire temperatures may alter soil mineral composition through thermal processes while removing vegetation, thus affecting soil spectral characteristics and leading to misclassification by NBR under certain complex surface conditions [43,44]. Supplementary indices such as MIR/Red have proven effective in separating combustion signals, but their applicability to high spatio-temporal resolution data still requires systematic verification [45]. For post-fire recovery assessment, long time series spectral trajectory analysis has become the mainstream method [46,47], and time series segmentation algorithms such as LandTrendr have enabled refined identification of recovery process turning points [48]. However, most existing studies focus on describing single recovery trends and rarely explore recovery spatial heterogeneity in depth with methods like the pixel dichotomy model [49].

Given the complex and multi-scale ecological impacts of wildfires, understanding how ecosystems respond to and recover from disturbances has become a central research focus [50]. Vegetation resilience has increasingly become a key concept for understanding ecosystem responses to wildfire disturbances. In ecological theory, resilience was first defined by Holling as “a measure of the ability of systems to absorb changes in state variables, driving variables and parameters and still exist” [51], referring to the capacity of ecosystems to absorb disturbances while maintaining their fundamental structure and functions. The concept has been further developed and widely applied to both social and ecological systems [52,53,54,55]. In a wildfire ecosystem, vegetation resilience is commonly reflected in the speed, trajectory, and patterns of post-fire vegetation regeneration [50,56] with post-fire recovery trajectories providing critical insights into ecosystem functioning and resilience [57]. Moreover, recovery patterns exhibit strong spatial variability driven by climatic, topographic, and biotic factors [58]. Vegetation resilience in this study denotes the capacity of burned vegetation to initiate regrowth and partially regain greenness within the first 180 days after fire disturbance. Rather than indicating full ecological recovery, it reflects the early adaptive response and regrowth potential of vegetation under post-fire environmental conditions. It should be noted that this definition represents an operational, NDVI-based interpretation of vegetation resilience in the context of this study, rather than a universally agreed conceptual equivalence. Accordingly, the dimensions of vulnerability, robustness, and recoverability are treated as analytical components derived from NDVI trajectories, rather than an exhaustive definition of resilience itself.

Most notably, the vast majority of existing studies have addressed burned area identification and post-fire recovery monitoring as two separate research topics, with few attempts to integrate them into a unified, end-to-end analytical framework. Furthermore, there remains a lack of systematic analysis that links fine-scale burned area delineation, quantitative characterization of recovery trajectory heterogeneity, and attribution analysis of driving factors, which hinders the comprehensive understanding of wildfire disturbance regimes and post-fire ecosystem responses. This is the core research gap this study aims to fill.

To address the deficiencies of existing studies in burned area identification accuracy, recovery heterogeneity characterization and driving mechanism analysis, this study aims to construct a high-precision framework for automatic wildfire extraction and post-fire recovery monitoring based on Sentinel-2 data.

(1)

Develop an automated high-precision burned-area delineation method using Sentinel-2 and optimized spectral indices (e.g., MIR/Red) to reduce boundary ambiguity and mixed-pixel effects.

(2)

Identify post-fire vegetation recovery trajectories through time-series clustering and quantify vegetation resilience patterns within burned areas.

(3)

Reveal wildfire recovery drivers by linking spatiotemporal recovery patterns with environmental factors such as topography, hydrothermal climate, and vegetation type.

Overall, the proposed framework advances wildfire monitoring from detection to process understanding, offering methodological innovation and quantitative evidence to support post-fire ecological restoration and management decision-making.

2. Materials and Methods

As shown in Figure 1, this study develops an integrated workflow for wildfire detection, recovery assessment, and mechanism attribution using multi-source data. Satellite imagery (Sentinel-2, Landsat-8, MODIS) and environmental datasets (CHIRPS precipitation, SRTM topography, Dynamic World land cover) are first harmonized through atmospheric correction, spatial resampling (30 m), and temporal alignment. Burned areas are then mapped using NDVI and MIR/Red indices to automatically delineate fine fire boundaries and reduce mixed-pixel effects, with accuracy validated against MODIS references. Vegetation resilience is quantified from NDVI time-series trajectories to identify disturbance and recovery phases. Finally, an interpretable machine-learning framework combining XGBoost and SHAP evaluates the importance and interactions of environmental drivers. The workflow links fire detection, recovery dynamics, and causal analysis within a unified framework for understanding post-fire ecosystem processes.

Geospatial data processing and computational analyses in this study were implemented using a suite of specialized software and libraries. Preprocessing of satellite imagery and calculation of spectral indices were performed on the Google Earth Engine platform. Landscape structure metrics were computed using Fragstats (version 4.3beta-64, University of Massachusetts, Amherst, MA, USA) [59]. The machine learning model and explainable attribution analysis were developed in Python (version 3.12.9), utilizing XGBoost (version 3.1.2) for regression modeling and SHAP (version 0.50.0) for feature interpretation.

2.1. Study Area and Data Collection

By synthesizing news reports, government bulletins, and satellite archives, five wildfire events in China were selected as research cases. The case site is referred to by the name of the city where it is located (Figure 2). These sites were chosen for their significant disaster scales, social impacts, and data availability. As detailed in

Table 1. Socio-economic impacts of the wildfire events.

Case Site	Start Date	End Date	Socio-Economic Impacts
Qinyuan	29 March 2019	2 April 2019	~942 ha burned; 24.7 k people affected; 6 fatalities recorded.
Yuanzhou	13 March 2021	14 March 2021	~267 ha burned; 2 fatalities and 6 injuries during suppression.
Xintian	17 October 2022	25 October 2022	Duration 8 days; ~5130 ha burned; 2 fatalities; $3.6 M (26 M CNY) loss.
Yuhuan	5 March 2022	8 March 2022	~8 ha burned; significant disruption to the local waxberry industry supply.
Yajiang	15 March 2024	28 March 2024	~5.33 ha burned; 3.4 k people affected; loss of high-value Matsutake habitats.

and

Table 2. Environmental characteristics of the case site.

Case Site	Coordinates	Climatic Type	Topography	Vegetation Type
Qinyuan	36.77° N, 112.30° E	Warm temperate continental monsoon climate	Dense forest with an average elevation of 1400 m.	Pinus tabuliformis
Yuanzhou	35.97° N, 106.10° E	Mid-temperate semi-arid continental monsoon climate	Loess Plateau landform (≈2000 m)	Wild grasses
Xintian	26.01° N, 112.19° E	Mid-subtropical continental monsoon climate	Southern foot of Yangming Mountain	subtropical evergreen broad-leaved forest
Yuhuan	28.27° N, 121.35° E	Subtropical monsoon climate (Maritime)	Coastal low mountains and hilly regions	Waxberry plantations
Yajiang	30.14° N, 101.13° E	Plateau Mountain climate	Average elevation of 2600 m	Quercus glauca and Pinus species

, the cases span a broad ecological gradient from 2019 to 2025 [60,61,62,63,64,65]. They range from the semi-arid loess highlands in Yuanzhou (elevation ≈ 2000 m) to the humid coastal hills in Yuhuan and the steep alpine valleys in Yajiang (avg. elevation 2600 m). Vegetation types also vary significantly, including Pinus tabuliformis in Qinyuan, deciduous forests in Xintian, and fruit plantations in Yuhuan. The diverse impacts, ranging from Xintian’s large-scale burn (5130 ha) to Yuhuan’s disruption of the waxberry industry, provide a solid basis for comparing post-fire recovery across different ecosystems. Their land use patterns also exhibit rich and complex characteristics. Accordingly, this study conducts a comparative analysis of post-fire vegetation recovery across these sites to examine differences among distinct ecological systems.

To analyze post-fire recovery and its driving factors, multi-source geospatial datasets were integrated and processed (

Table 3. Summary of multi-source remote sensing datasets.

Dataset	Source	Resolution	Revisit Period
Harmonized Sentinel-2 MSI (Level-2A SR)	European Space Agency	10–20 m	5 days
MOD11A1.061 (LST and Emissivity)	U.S. Geological Survey	1200 m	1 day
CHIRPS Daily (Precipitation, Ver. 2.0)	UC Santa Barbara CHC	5566 m	1 day
USGS Landsat 8 (Level 2, Collection 2)	U.S. Geological Survey	30 m	16 days
NASA SRTM Digital Elevation Model	NASA	30 m	–
Dynamic World V1 (Land Use)	World Resources Institute	10 m	2–5 days

). Harmonized Sentinel-2 MSI (Level-2A) imagery, with a 10–20 m resolution and a 5-day revisit cycle, was filtered to have both the proportion of pixels with the Normalized Difference Snow Index (NDSI) < 0.4 and the GEE-determined CLOUDY_PIXEL_PERCENTAGE less than 10%, serving as the primary data source for calculating the Normalized Burn Ratio (NBR), MIR/Red and NDVI. To supplement the analysis of surface environmental conditions, MOD11A1 and Landsat 8 (Level 2) products were utilized to derive Land Surface Temperature (LST). Precipitation data were extracted from the CHIRPS Daily dataset (Version 2.0) at a resolution of 5566 m. For topographic analysis, the NASA SRTM Digital Elevation Model (30 m) was employed to calculate elevation, slope, and aspect. Additionally, land use types were identified using the Dynamic World V1 dataset (10 m resolution). All data processing, including atmospheric correction and index calculations, was conducted to ensure temporal and spatial consistency across the five study sites, and the data scope covered the 3 years before and 2 years after the wildfire occurrence.

2.2. Fire Sentinel-2–Based Wildfire Extraction and Burned Area Identification

The Ratio Vegetation Index (RVI), also referred to as the Simple Ratio (SR) Index and computed from the NIR/Red band ratio, was first proposed by Jordan in 1969. Based on the spectral characteristics of vegetation leaves, this index enables the effective extraction of forest features by utilizing the ratio of infrared to red bands in remote sensing applications [66] and has been extensively adopted in vegetation monitoring research [67]. The MIR/Red Index, a derivative of the RVI, has demonstrated excellent efficacy in extracting vegetation information typified by forest features [68] and has also been applied to aerosol retrieval [69]. Given that high temperatures induced by wildfires can alter soil mineral components and thereby modify their spectral characteristics while damaging vegetation, and that the mid-infrared (MIR) band is particularly suitable for monitoring high-temperature targets and minerals in comparison with the near-infrared (NIR) band, this index has also exhibited outstanding performance in hot spot and wildfire detection [70].

The Normalized Difference Vegetation Index (NDVI) was proposed by Rouse in 1973 for the identification of vegetation features [71]. Owing to its robust and superior extraction performance, it has been widely applied in wildfire extraction studies where vegetation loss is the primary characteristic [72].

In this study, drawing on the spectral band characteristics of the Sentinel-2 satellite, the MIR/Red Index was constructed by selecting its Red Edge Band 1 (B5) and Shortwave Infrared Band 2 (B12). In terms of the calculation methodology, remote sensing image data acquired within 10 days prior to the wildfire occurrence were selected and subjected to best pixel composite processing, which served as the baseline for evaluating the pre-fire ecological status of the study area. Meanwhile, imagery collected within 3 to 13 days after the wildfire event was processed following the identical procedure.

Subsequently, a comparative analysis was performed between the pre-fire and post-fire images. For the first thresholding step, pixels were screened with the following criteria: the NDVI difference (calculated as pre-fire NDVI minus post-fire NDVI) greater than 0.1, a pre-fire NDVI value higher than 0.16, and a post-fire NDVI value lower than 0.14. For the second thresholding step, pixels were retained when meeting both requirements: the MIR/Red Index difference (calculated as post-fire MIR/Red Index minus pre-fire MIR/Red Index) greater than 0.5, and a post-fire MIR/Red Index value higher than 1. By combining the two-step thresholding criteria, the MIR/Red Index and NDVI jointly enable accurate extraction of wildfire-affected areas.

2.3. NDVI-Driven Vegetation Resilience Quantification and Post-Fire Recovery Dynamics

The resilience V-shaped curve refers to a conceptual representation that depicts the temporal response of a system’s performance following an external disturbance, typically characterized by an abrupt decline and a subsequent recovery toward a relatively stable state [73,74,75]. This curve has been widely applied in resilience analysis to examine how systems or networks absorb shocks and restore functionality after disturbance events [76]. The conceptual foundation of the V-shaped resilience framework can be traced back to ecological resilience theory, which emphasizes the capacity of a system to withstand disturbances while maintaining its fundamental structure and core functions [51]. In recent years, its underlying decline–recovery logic has been increasingly adopted in vegetation and disturbance ecology. In remote sensing-based post-fire studies, this framework is commonly operationalized through NDVI and other vegetation index time series by quantifying disturbance-induced decline, recovery rate, and recovery extent, thereby characterizing post-fire vegetation response trajectories and ecosystem resilience [50,77].

In this study, post-fire NDVI trajectories were analyzed using a V-shaped response framework to quantify immediate damage, recovery speed, and recovery completeness across multiple fire-affected regions. As shown in Figure 3, the V-shaped NDVI resilience curve is divided into four stages: Original State (pre-disruption), Phase I (decline), Phase II (trough), and Phase III (recovery), which correspond to three key resilience dimensions, namely vulnerability, robustness, and recoverability, respectively [78,79].

2.3.1. Original State

The Original State represents pre-fire vegetation conditions ($NDV{I}_{pre}$), during which NDVI typically exhibits pronounced seasonal fluctuations driven by phenological cycles and climatic variability. Remote sensing studies have shown that in undisturbed ecosystems, NDVI time series inherently reflect vegetation growth cycles and seasonal climate signals [80,81,82], supported by observed spatiotemporal variation patterns linked to climatic drivers such as temperature and precipitation [83].

To ensure the robustness of the baseline representation, we further evaluated NDVI variability over different pre-fire temporal windows (one-year, two-year, and three-year periods). The comparison results indicate that interannual NDVI variations in the study area are relatively small. This suggests that the pre-fire vegetation state is stable and not sensitive to short-term climate variability.

To quantify the stability of this baseline, the coefficient of variation ($C{V}_{pre}$) was calculated during the one-year pre-fire period:

$$C{V}_{pre}={\displaystyle \frac{{\sigma }_{pre}}{{\mu }_{pre}}}$$

(1)

where ${\sigma }_{pre}$ is the standard deviation and ${\mu }_{pre}$ is the mean NDVI of the pre-fire year. A low $C{V}_{pre}$ ensures that the $NDV{I}_{pre}$ (the arithmetic mean of this period) serves as a robust and representative benchmark for subsequent resilience analysis. The wildfire occurrence date was defined as $t_0$.

2.3.2. Phase I (Vulnerability)

Begins at $t_0$ and ends at $t_1$ (the lowest point), corresponding to the NDVI nadir. This phase captures the system’s immediate response to wildfire disturbance. A longer duration and a greater decline in magnitude (e.g., reflecting higher burn severity) indicate higher sensitivity and exposure of vegetation to wildfire disturbance. Vulnerability, representing the sensitivity of vegetation to wildfire disturbance, was quantified using three complementary indicators: the NDVI nadir ($NDV{I}_{min}$), the relative drop magnitude (drop (%)), and interannual productivity change ($\mathrm{\Delta }NDVI$).

The NDVI nadir corresponds to the minimum NDVI value observed after the fire event and reflects the maximum instantaneous vegetation damage. The proportional reduction in NDVI relative to the pre-fire baseline was calculated as:

$$Drop (\%)={\displaystyle \frac{NDV{I}_{pre}-NDV{I}_{min}}{NDV{I}_{pre}}}\times 100$$

(2)

Interannual productivity change was calculated as:

$$\mathrm{\Delta }NDVI=NDV{I}_{pre}-NDV{I}_{fire}$$

(3)

where $NDV{I}_{fire}$ denotes the mean NDVI in the fire year. A negative ΔNDVI indicates that annual vegetation productivity exceeded pre-fire levels, suggesting low vulnerability or rapid compensation effects.

2.3.3. Phase II (Robustness)

From $t_1$ to $t_2$, where values remain within ±30% of the maximum drop magnitude for at least two months. If this phase is absent, $t_1=t_2$ and robustness = 0. This phase reflects the system’s capacity to maintain residual functioning at the lowest performance level before recovery initiates.

To quantify robustness, the trough duration ($D_{trough}$) was quantified in months as:

$$D_{trough }\left(M\right)=t_2-t_1$$

(4)

While a shorter or absent trough implies a rapid transition to recovery, a stable $D_{trough}$ indicates the system’s ability to withstand further degradation under post-fire environmental stress, characterizing the buffering capacity of the vegetation community before active regrowth begins.

2.3.4. Phase III (Recoverability)

It starts after $t_2$ (or $t_1$ if no trough) and concludes at ${t_0}^{\prime }$, the point at which NDVI returns to the pre-fire baseline level. This phase represents the active and final stages of the ecosystem returning to pre-shock operational levels; a longer duration in this phase indicates a more protracted restoration of vegetation productivity. Recoverability was evaluated using recovery duration ($D_{rec}$), recovery rate ($R_{rate}$), and recovery extent ($R_{extent}$), to capture the speed, efficiency, and completeness of the recovery process. To enable comparison across sites with different fire years and observation windows, $D_{rec}$ and $R_{rate}$ were calculated relative to each site’s pre-fire NDVI baseline. This standardization allows assessment of relative recovery dynamics while acknowledging that the absolute post-fire timelines may differ due to environmental and temporal factors.

The recovery duration ($D_{rec}$), measured in months, quantifies the total time required for the ecosystem to revert to its pre-disturbance baseline:

$$D_{rec}\left(M\right)={t_0}^{\prime }-t_0$$

(5)

A shorter $D_{rec}$ signifies a more efficient and rapid restoration of ecological functions.

The recovery rate ($R_{rate}$) was defined as the linear slope of NDVI increase from the nadir to the end of the first post-fire year:

$$R_{rate}={\displaystyle \frac{NDV{I}_{y1}-NDV{I}_{min}}{n}}$$

(6)

where $NDV{I}_{y1}$ is the mean NDVI during the first post-fire year and $n$ denotes the number of observation intervals.

The recovery extent ($R_{extent}$) was calculated as:

$$R_{extent} (\%)={\displaystyle \frac{NDV{I}_{y2}}{NDV{I}_{pre}}}\times 100$$

(7)

where $NDV{I}_{y2}$ is the mean NDVI during the second post-fire year.

2.4. Explainable Machine Learning–Based Contribution Analysis of Fire Resilience Drivers

2.4.1. Extreme Gradient Boosting Model

We employed an XGBoost regression model to quantify the nonlinear effects of multiple environmental drivers on vegetation recovery after fire. To ensure modeling rigor, predictors were pre-screened using Variance Inflation Factors (VIF) to eliminate multicollinearity. This structure captures complex nonlinear relationships and higher-order interactions in high-dimensional feature spaces, making it well-suited for ecological remote sensing data characterized by strong spatial heterogeneity. This structure captures complex nonlinear relationships and higher-order interactions in high-dimensional feature spaces and is well-suited to ecological remote sensing data characterized by strong spatial heterogeneity, as it can effectively internalize spatial dependencies into explained variance. Recent studies have demonstrated the high performance of XGBoost in modeling forest structural recovery [84] and identifying optimal restoration potential across complex landscapes [85].

To ensure spatiotemporal comparability and prevent overfitting, a consistent and robust set of hyperparameters (Learning Rate = 0.05, Max Depth = 6, N Estimators = 200, Tree Method = hist) was applied across all models. This fixed-parameter strategy ensures that variations in feature importance reflect actual environmental mechanisms rather than fluctuations in model complexity or artifacts of site-specific parameter tuning.

2.4.2. SHAP-Based Model Interpretation

To enhance interpretability, we decomposed model predictions using the SHAP (SHapley Additive exPlanations) framework [86]. The SHAP framework [86] was utilized to decompose model predictions into local, pixel-level attributions, effectively capturing the complex spatial heterogeneity of post-fire recovery. For any pixel sample, the decomposition can be written as:

$$g\left(Z^{\prime }\right)={\phi }_0+\sum _{j=1}^M{\phi }_j{Z^{\prime }}_j\approx f\left(x\right)$$

(8)

where $g$ is the explanation model, $f\left(x\right)$ is the original predictive model, $Z^{\prime }={\left\{0, 1\right\}}^M$ encodes feature inclusion, $M$ is the total number of features, ${\phi }_j$ denotes the attribution for feature $j$, and ${\phi }_0$ is the baseline. The attribution for feature $j$ is given by the Shapley value:

$${\phi }_j\left(f\right)=\sum _{S\subseteq N_s/\left\{i\right\}}{\displaystyle \frac{\left|S\right|!\left(M-\left|S\right|-1\right)!}{\left|N_S\right|!}}\left[f\left(S\bigcup \left\{i\right\}\right)-f\left(S\right)\right]$$

(9)

where ${\phi }_j\left(f\right)$ denotes the contribution of feature $j$ to the original model function $f$, $N_S$ is the set of all features and $S$ denotes a subset, $\left|S\right|$ is its cardinality, and $f\left(S\right)$ is the model prediction when only features in $S$ are present.

This formulation computes the weighted average marginal contribution of feature $j$ across all possible subsets, thereby quantifying its independent effect on the prediction. For tree-based models, efficient approximation algorithms enable scalable computation of Shapley values at the pixel scale. Consequently, the recovery increment of each pixel can be expressed as the sum of contributions from different drivers, providing a foundation for threshold detection and interaction analysis.

2.4.3. Variable Definition and Data Processing

All 30 m resolution pixels within each burned area were treated as modeling units. The dependent variable is the net vegetation recovery increment ΔNDVI, which removes pre-fire differences in background vegetation conditions and isolates post-fire regenerative capacity.

Predictors were grouped into three categories: climate, topography, and landscape configuration. Climate variables include cumulative precipitation and cumulative land surface temperature, calculated as daily sums from the ignition date to the observation date to represent total water input and accumulated thermal exposure during the recovery period. Cumulative precipitation reflects potential water availability, whereas cumulative land surface temperature characterizes thermal conditions and potential heat stress. Topographic variables include elevation, slope, and aspect. Because the aspect is circular, it was transformed into sine and cosine components to represent a continuous radiation gradient and avoid information loss associated with categorical classification. Landscape configuration metrics were derived from 10 m resolution land use data. The Shannon diversity index was computed using Fragstats within a 100 m radius moving window to quantify habitat heterogeneity and spatial structural complexity. All predictors were harmonized to 30 m resolution, and anomalous values were removed to ensure stability and comparability.

2.4.4. Analytical Framework

The analytical procedure comprised three components. First, SHAP attributions were across three specific recovery stages (20, 90, and 180 days post-fire) to evaluate temporal shifts in the relative importance of drivers and to identify stage-specific controlling factors. Second, SHAP dependence plots were examined to characterize response curves between key predictors and ΔNDVI, detect potential threshold intervals, and identify nonlinear inflection points. Third, we assessed how topographic context modulates the effects of climate and landscape variables, thereby revealing synergistic and constraining interactions among environmental drivers during vegetation recovery after fire.

By integrating an ensemble learning algorithm with an interpretable attribution framework, this approach improves predictive performance while enabling pixel-scale tracing of driver contributions, thereby supporting a more mechanistic analysis of ecological recovery in spatially heterogeneous landscapes.

3. Results

3.1. Spatial Distribution of Wildfire Occurrence and Performance of Sentinel-2–Based Extraction

To validate the accuracy of the self-developed burned area extraction approach, this study conducted a horizontal comparison with two existing MODIS-based fire products and adopted random sampling points to perform the accuracy assessment simultaneously. Specifically, random points were generated within the study areas using a random number function. For each individual study area, 50 fire pixels and 50 non-fire pixels were then labeled via manual visual interpretation (Figure 4, column 1), based on which the subsequent accuracy assessment was carried out.

The accuracy assessment based on stratified random sampling demonstrates that the burned area extracted from Sentinel-2 imagery significantly outperforms MODIS-based fire products. Overall accuracy of the Sentinel-2 burned area map reaches 99.8%, compared to 77.9% and 92.7% for the FIRMS product and MCD64A1 product separately. In terms of class-specific performance, since existing products mostly exhibit larger and continuous coverage areas as well as coarser resolution, all three methods deliver excellent performance with their producer’s accuracies approaching 100%. On the contrary, the user’s accuracy for burned areas derived from Sentinel-2 is 98.9%, substantially exceeding that of FIRMS (76.3%) and MCD64A1 (86.3%), suggesting fewer commission errors. These results indicate that the higher spatial resolution of Sentinel-2 enables more precise delineation of fire boundaries and small or fragmented burned patches that are often underestimated or omitted in coarser-resolution MODIS products.

The proposed burned area extraction method based on Sentinel-2 imagery was further applied to five representative fire-prone regions in China, including Qinyuan, Xintian, Yuanzhou, Yuhuan and Yajiang (Figure 4). The total burned areas identified by Sentinel-2 were estimated at 91.8 km², 38.4 km², 6.9 km², 3.9 km², and 186.4 km², respectively. Expressed in such a way that the first percent value denotes the overestimation extent of the FIRMS product compared with the results independently extracted in this study, and the second percent correspondingly represents that of the MCD64A1 product (FIRMS/MCD64A1), the overestimation of the case study areas can be presented as follows: MODIS-based products consistently report larger burned areas, with overestimation of approximately 208.8%/39.8% in Qinyuan, 262.8%/173.2% in Xintian, 352.2% in Yuanzhou, 561.5%/141.0% in Yuhuan, and 68.6%/41.6% in Yajiang relative to Sentinel-2 results. The overestimation mentioned above is speculated to be mainly manifested in the fact that the two MODIS-based fire products have spatial resolutions of 500 m (MCD64A1) and 1000 m (FIRMS), respectively, which are far coarser than the 10 m resolution used in this study for burned area extraction based on Sentinel-2. In addition, the burned area product of FIRMS is produced by first detecting fire centroid pixels at a 1000 m scale and then extending burned area coverage around them at the same resolution, which is more indicative of “fire point detection” rather than “regional burned area extraction”. As a result, both products show a tendency to overestimate in burned area estimation. The largest differences are observed in Yuanzhou and Yuhuan, China, where complex terrain and fragmented fire patterns prevail. Spatial inspection indicates that Sentinel-2 captures finer fire boundaries and small, discontinuous burned patches that are frequently omitted or generalized in coarser-resolution products. Moreover, the existing products show greater extraction errors in cases of wildfires with a smaller total area.

These results demonstrate that integrating the MIR/Red Index and NDVI via two-step thresholding, along with Sentinel-2’s 10 m fine resolution, enables accurate, stable burned area extraction. It captures small fragmented fire patches and clearly outperforms coarse-resolution MODIS fire products.

3.2. Post-Fire Vegetation Resilience Trajectories Revealed by NDVI Dynamics

Figure 5 illustrates distinct NDVI recovery trajectory patterns across the study sites. Three typical forms can be identified: V-shaped trajectories (Qinyuan, Xintian, and Yuhuan), an asymmetric stepwise recovery pattern (Yuanzhou), and a wide V-shaped trajectory with delayed recovery (Yajiang).

According to

Table 4. Quantitative metrics of NDVI trajectory characteristics across case sites.

Phase	Site	(a) Qinyuan	(b) Yuanzhou	(c) Xintian	(d) Yuhuan	(e) Yajiang
Original State	$C{V}_{pre}$	0.476	0.764	0.265	0.421	0.211
Vulnerability	$NDV{I}_{min}$	0.017	0.068	0.002	0.040	0.052
	$\mathrm{\Delta }NDVI$	0.205	−0.046	0.236	0.166	0.362
	$Drop (\%)$	96.00	76.90	99.55	90.34	89.58
Robustness	$D_{trough}\left(M\right)$	1.3	0.7	1.3	3.3	-
Recoverability	$D_{rec}$(M)	3.2	1.8	11.2	7.2	-
	$R_{rate}$	0.025	0.0239	0.0357	0.0239	-
	$R_{extent }(\%)$	68.90	128.78	78.27	117.21	47.6

, the three core dimensions of resilience exhibit significant spatial heterogeneity. All regions displayed distinct seasonal NDVI cycles prior to the wildfire. Pre-fire coefficient of variation values ranged from 0.211 to 0.764. Yajiang and Xintian maintained the highest stability. Yuanzhou exhibited the greatest volatility. To better interpret the magnitude of disturbance across sites, burn severity was further characterized using dNBR and RdNBR indices (Figure 6). These indices provide complementary information on fire-induced vegetation damage and help contextualize NDVI-based responses.

In the decline phase, all sites reached their NDVI nadir within 3 to 6 months post-fire. Xintian suffered the most severe immediate damage. Its NDVI dropped by 99.55% to a minimum of 0.002. Yuanzhou recorded the smallest NDVI decline at 76.90%. Its fire-year NDVI stayed above the pre-fire average (ΔNDVI = −0.046). No persistent ecological degradation occurred in Yuanzhou during the disturbance year. Conversely, Yajiang sustained the most substantial vegetation loss (ΔNDVI = 0.362), far exceeding other regions and indicating acute functional vulnerability. This pattern is consistent with burn severity indicators, as Yajiang exhibits the highest dNBR and RdNBR values among all sites, suggesting intense fire disturbance. Similarly, Xintian also shows relatively high burn severity, which corresponds to its pronounced NDVI decline. In contrast, Qinyuan displays comparatively lower burn severity, aligning with its more moderate vegetation loss. These results suggest that differences in initial disturbance intensity partly explain the observed variability in NDVI decline across sites.

In the trough phase, Yuanzhou had the shortest duration (0.7 months), followed by Qinyuan and Xintian (1.3 months). Yuhuan experienced the longest trough (3.3 months), and vegetation functions were heavily impacted. However, its minimum NDVI remained above historical extremes. The ecosystem thus maintained basic operational capacity. Notably, Yajiang exhibited a continuous low-level NDVI. This resulted from its recent fire date and persistent suppression.

In the recovery phase, recovery efficiency varied significantly across regions. The time required for NDVI to reach pre-fire levels followed a specific order. These were Yuanzhou (1.8 months), Qinyuan (3.2 months), Yuhuan (7.2 months), and Xintian (11.2 months). Xintian achieved the highest recovery rate ($R_{rate}$ = 0.0357). Other regions averaged approximately 0.023. For Yajiang, an effective recovery trajectory has not yet been fully observed; consequently, its recovery duration and rate are currently not quantified. Its relatively high burn severity further suggests that both strong disturbance intensity and limited observation time contribute to its apparent low resilience. Regarding the recovery extent, only Yuanzhou and Yuhuan achieved compensatory growth. Their NDVI values exceeded the baseline. The remaining three regions did not reach 80% of their historical averages. Specifically, Yajiang stayed below the 50% threshold. This represents severe ecological vulnerability.

The analysis reveals significant spatial heterogeneity in vegetation responses to wildfire disturbances. This study identifies three distinct trajectories of post-fire resilience. It should be noted that these patterns reflect relative recovery performance under differing levels of initial burn severity. (1) Medium Resilience: Exemplified by Yuanzhou, this pattern features a stepwise increase in NDVI. It is characterized by a synergistic combination of low vulnerability and robust recoverability, indicating high systemic stability. Its moderate burn severity further supports a relatively balanced disturbance–recovery dynamic. (2) High Resilience: Exemplified by Yuhuan, this mode exhibits exceptional restoration kinetics, with both recovery efficiency and rate reaching peak levels. It is characterized by superior recoverability and overcompensation, suggesting strong regenerative capacity despite moderate fire disturbance. (3) Low Resilience: Exemplified by Qinyuan, Xintian, and Yajiang, this pattern is defined by compromised post-fire regenerative capacity. These sites collectively exhibit heightened ecological vulnerability and a protracted recovery lag. In particular, higher burn severity in Xintian and Yajiang contributes to their reduced recovery performance, while Qinyuan reflects relatively lower disturbance but still limited recovery efficiency.

Overall, NDVI dynamics reveal distinct post-fire vegetation resilience trajectories across the study sites, highlighting pronounced spatial heterogeneity in both recovery patterns and resilience capacities.

3.3. Key Drivers and Relative Contributions to Vegetation Resilience Identified by Explainable Models

3.3.1. Applicability of the XGBoost Model

The study developed an XGBoost regression model to capture the nonlinear relationships between post-wildfire vegetation recovery increment (ΔNDVI) and multiple environmental factors. Model performance was evaluated using the test R² and Mean Absolute Error (MAE) based on an independent 80/20 data split (

Table 5. XGBoost model performance evaluation and spatial autocorrelation diagnostics across case sites and time intervals.

Location	Time	Test R2	MAE	Moran Residual	Moran Original
Qinyuan	20 days	0.528	0.036	0.547	0.796
	90 days	0.474	0.094	0.563	0.783
	180 days	0.478	0.083	0.568	0.782
Xintian	20 days	0.335	0.049	0.473	0.623
	90 days	0.421	0.042	0.398	0.620
	180 days	0.443	0.050	0.460	0.665
Yajiang	20 days	0.752	0.036	0.561	0.882
	90 days	0.740	0.056	0.606	0.884
	180 days	0.624	0.092	0.644	0.859
Yuanzhou	20 days	0.385	0.013	0.347	0.635
	90 days	0.814	0.062	0.607	0.957
	180 days	0.629	0.032	0.481	0.806
Yuhuan	20 days	0.249	0.022	0.266	0.465
	90 days	0.571	0.038	0.473	0.793
	180 days	0.684	0.051	0.525	0.849

). Across five representative case sites and multiple post-fire time intervals, the R² values of the models range from 0.25 to 0.81 (detailed in Table 5). These values demonstrate a consistent and strong correspondence between the predicted and observed recovery patterns across varying ecological contexts.

From a temporal perspective, model performance exhibits clear stage-dependent characteristics. As shown in Figure 7, during the early recovery period (20–90 days post-fire), model performance is generally strong across most study areas. For instance, at 90 days post-fire, the R² values for Yuanzhou and Yajiang reach 0.81 and 0.74, respectively. In contrast, lower explanatory power is observed in certain regions during the initial recovery stage; for example, Yuhuan shows a relatively low R² of 0.25 at 20 days post-fire, highlighting inter-site differences in spatial variability during early recovery. When the recovery period extends to 180 days, model explanatory power declines in some study areas but remains within an acceptable range overall, indicating that environmental factors continue to exert a sustained influence on the spatial variation in vegetation recovery over longer time scales. Notably, spatial diagnostic tests (Table 5) reveal a significant reduction in the Global Moran’s I from the original ΔNDVI (0.46–0.96) to the model residuals (0.27–0.64). This substantial decrease in spatial clustering demonstrates that the model successfully internalized the spatial structure of the landscape into biophysical attributions, rather than relying on spatial proximity bias.

3.3.2. Relative Importance of Driving Factors

The relative importance results calculated based on SHAP values show that post-fire vegetation recovery in the five case sites exhibits clear temporal succession characteristics. However, significant differences exist in the dominant factor structures and their evolutionary paths across regions, reflecting differentiated recovery patterns under different geographical backgrounds.

The ΔNDVI recovery type in the Qinyuan region shows a transition from climate dominance to topographic control. During the 20–90 days post-fire stage, the combined proportion of cumulative precipitation and cumulative land surface temperature remains stable, the contribution rate of the landscape diversity index SHDI is maintained at approximately 25–27%, and elevation accounts for about 23–26%. The recovery process in this stage is mainly influenced jointly by hydrothermal conditions and patch structure. After entering 180 days, the importance of elevation increases to 33.2%, becoming the primary driving factor, and the slope aspect factor, Aspect (Cos), increases to 16.0%, while SHDI decreases to 12.4%. This change indicates that as vegetation communities gradually establish, the promoting role of landscape structure weakens, whereas the radiation conditions and moisture distribution patterns shaped by topography gradually occupy a dominant position.

The Xintian region presents recovery characteristics jointly dominated by thermal environment and topography. Within 20 days after the fire, the contribution rate of cumulative LST reaches 32.2%, and elevation accounts for 27.2%. The temperature factor occupies a dominant position in the initial stage. During the 90–180 day stage, elevation remains stable at about 30%, Aspect (Cos) rises to about 20.6%, and SHDI consistently remains below 10%. This structural change indicates that the recovery process in this region is relatively sensitive to thermal conditions. With time, the regulating effect of slope aspect on radiation input strengthens, while the explanatory power of landscape pattern on recovery increment remains limited throughout.

The Yajiang region exhibits a high-mountain recovery pattern dominated by moisture and slope aspect over the long term. Cumulative precipitation accounts for a relatively high proportion at all stages, being 36.2% at 20 days post-fire and remaining at 24.5% at 180 days post-fire. Meanwhile, Aspect (Cos) increases from 24.3% to 34.4%. In contrast, SHDI remains at a low level of 5–8%. These results reflect that in high-altitude areas, water supply plays a key role in the early stage of recovery. As recovery progresses, the radiation differences determined by slope aspect gradually exert a stronger influence on vegetation growth.

The driving structure in the Yuanzhou region shows stage-based changes. At 20 days post-fire, cumulative precipitation accounts for 27.3% and LST for 22.7%, jointly constituting the main driving forces. At the 90-day stage, the contribution rate of LST rises to 43.9%, reaching the highest value over the entire period, indicating that mid-term recovery is highly sensitive to thermal conditions. By 180 days post-fire, the contribution of climatic factors decreases, while Aspect (Cos) increases to 19.8%. The dominant factors of the recovery process shift from climatic conditions to microhabitat differences regulated by slope aspect. SHDI remains below 9% at all stages, indicating that the explanatory strength of landscape structure for recovery in this area is relatively limited.

The Yuhuan region shows coastal hilly characteristics with a transition from responses to thermal conditions and elevation toward slope-aspect control. At 20 days post-fire, LST accounts for 31.4%, elevation for 23.3%, and SHDI for 13.7%. The initial recovery is jointly influenced by hydrothermal conditions and landscape structure. By 180 days post-fire, the combined proportion of slope aspect factors approaches 40%, becoming the most important driving factor, while the contribution of SHDI declines significantly. This change indicates that in topographically fragmented coastal hilly areas, medium- and long-term recovery increments are mainly influenced by local radiation and moisture conditions regulated by slope aspect.

Overall, based on the selected environmental predictors, different regions exhibit a general trend of shifting from climate dominance to topographic dominance during post-fire recovery, but the speed of transition and dominant structures differ significantly. High-altitude regions remain controlled by moisture and slope aspect over the long term, whereas hilly and coastal areas display more pronounced stage-specific thermal response characteristics. The explanatory power of landscape structure factors decreases over time in most regions, suggesting that their role within the current modeling framework is concentrated primarily in the early recovery stage.

3.3.3. Nonlinear Thresholds and Interaction Mechanisms of Driving Factors

SHAP dependence plots (Figure 8, Figure 9 and Figure 10, arranged with environmental features as rows and case sites as columns) and global interaction heatmaps show that the contributions of individual drivers to the vegetation recovery increment are generally nonlinear. Clear regional differences appear in response curve shapes, threshold locations, and interaction structures. These differences reflect distinct sensitivity windows to resource and structural conditions across habitat contexts rather than simple variation in effect magnitude.

In heterogeneous habitat regions (Qinyuan and Yuhuan), SHDI exhibits a clear sigmoidal response. When SHDI < 0.35, SHAP values remain near zero, indicating limited influence of low landscape heterogeneity on recovery. Between 0.35 and 0.75, SHAP values increase rapidly, suggesting that moderate heterogeneity substantially enhances recolonization through improved patch connectivity. When SHDI > 0.75, the response saturates and slightly declines, implying that excessive fragmentation reduces effective connectivity. Thus, the effect of SHDI is threshold-dependent rather than monotonic.

Elevation modulates this relationship. For the same SHDI, lower elevations show higher SHAP contributions than higher elevations. This stratification indicates that favorable hydrothermal conditions enhance the translation of structural heterogeneity into biomass recovery, whereas colder and thinner soils at high elevations limit recovery efficiency.

In thermally sensitive regions (Xintian and Yuanzhou), cumulative temperature follows a unimodal response. Recovery increases with temperature at low levels but declines beyond a threshold, indicating an optimal thermal window. High precipitation moderates this decline, suggesting interactive regulation between heat and moisture.

In cold and arid high-elevation regions (Yajiang), cumulative precipitation shows a logarithmic response. Small rainfall increases strongly promote early recovery, while marginal effects diminish after basic water demand is satisfied. Slope aspect further regulates recovery: higher cosine values correspond to stronger recovery contributions and lower surface temperature, reflecting improved moisture retention.

Overall, major drivers display nonlinear and threshold behavior. Landscape diversity is most effective within intermediate ranges, temperature exhibits an upper limit, and precipitation shows diminishing returns. Topography acts as a regulator by redistributing water and heat rather than as an independent driver. These interactions and threshold effects explain the divergence of post-fire recovery trajectories among regions.

4. Discussion

4.1. Spatial Organization of Post-Fire Resilience

The high Global Moran’s I values for ΔNDVI (0.46–0.96) confirm that post-fire recovery is a spatially structured process rather than a stochastic pixel-level response. This spatial clustering reflects the expansion from biological legacies such as unburned patches and seed banks. Similar neighborhood-controlled recovery has been reported in many fire-prone ecosystems, such as the boreal forests of North America and Siberia, the Greater Khingan Range, the California chaparral, and the tropical montane forests of Vietnam [50,87,88,89,90].

Notably, our results reveal that landscape heterogeneity (SHDI) also exhibits exceptionally high spatial autocorrelation (Moran’s I > 0.90), providing a robust structural foundation for the high functional contributions identified in the SHAP analysis. Such strong spatial dependency in the landscape configuration ensures that the availability of propagule sources and microclimatic refuges is systemically organized across the fire scar [91,92].

By incorporating these highly spatially structured predictors, the XGBoost model effectively internalized the inherent spatial dependency into biophysical explained variance. This is evidenced by the marked reduction in Moran’s I within the model residuals (decreasing to 0.27–0.64) compared to the original ΔNDVI. This demonstrates that recovery behaves as a connectivity-driven diffusion process, where the explanatory power is rooted in the biophysical drivers rather than simple spatial proximity artifacts [93]. Ignoring this spatial structure would not only overestimate climatic effects but also underestimate the critical role of structural ecological memory in buffering post-fire disturbances.

4.2. Temporal Succession of Dominant Recovery Drivers

The temporal decline in model explanatory power reflects a shift in regulatory mechanisms during succession. Immediately after fire, recovery is constrained primarily by abiotic factors, whereas later stages increasingly reflect endogenous ecosystem processes.

Within the first 20–90 days, accumulated precipitation and surface temperature have a dominant influence among the evaluated abiotic factors. This period corresponds to the post-disturbance establishment window, during which seedlings and resprouts encounter physiological survival thresholds [94]. Fire removes canopy buffering and exposes soil to intense radiation and evaporation, making water–energy balance a critical limiting factor [95]. At this stage, environmental filtering appears to strongly regulate whether regeneration can initiate, though other unquantified factors, such as initial burn severity, also play a fundamental role.

As recovery proceeds, the explanatory power of environmental predictors declines. This shift does not imply weaker environmental influence but rather indicates that community assembly processes, such as competition, soil development, and niche differentiation, become increasingly important [95,96]. The system thus transitions from externally constrained dynamics to self-organized ecological regulation, consistent with resilience theory [51]. Post-fire resilience can therefore be conceptualized as a two-phase system: climate-controlled regeneration initiation [97] and biotically regulated community reorganization [98].

4.3. Nonlinear and Interaction Mechanisms of NDVI Recovery

The nonlinear relationship between landscape diversity and vegetation recovery in this study indicates the existence of an optimal heterogeneity range. Moderate heterogeneity enhances recolonization by providing diverse niches and seed sources, whereas excessive fragmentation reduces habitat continuity and dispersal efficiency. This pattern agrees with the intermediate heterogeneity hypothesis and fragmentation theory [99].

Topography (elevation and aspect) further modulates this relationship by regulating microclimatic buffering capacity [100]. Moist and shaded slopes amplify the positive effect of landscape diversity by stabilizing temperature and water availability, while warm exposed slopes increase evaporative stress and weaken landscape effects. Similar terrain-mediated recovery sensitivity has been documented in mountainous fire systems, such as the foothill forests of southeastern Australia, the montane cordillera of western Canada, the boreal pine forests of southern Siberia, the fynbos biomes of South Africa, and the Mediterranean pine forests of central Spain [48,88,100,101,102].

These findings indicate that recovery drivers interact multiplicatively rather than additively: climate determines physiological thresholds, topography regulates microclimate buffering, and landscape structure controls dispersal and recolonization. Consequently, linear attribution models may oversimplify resilience dynamics, especially under changing climate conditions.

4.4. Regional Recovery Archetypes and Management Implications

The synthesis of driver importance and temporal dynamics reveals three distinct recovery archetypes in this study. The first one is the habitat-heterogeneity-facilitated recovery. Regions with moderate fragmentation recover rapidly due to preserved seed sources and microhabitats. Protection of unburned patches is more effective than large-scale replanting, consistent with the refuge-driven recovery paradigm [103].

The second one is the physical-constraint-limited recovery. In environmentally harsh regions, water and energy constraints dominate throughout succession. Natural recovery remains spatially restricted, indicating the necessity of assisted restoration or soil stabilization. Comparable regeneration failures have been observed in climatically stressful ecosystems [104].

The third one is the thermal-threshold-sensitive recovery. In warm climates, regeneration depends on a short thermal window during early establishment. Small temperature anomalies strongly influence survival probability, suggesting that restoration timing is more critical than restoration intensity [105]. Together, these archetypes demonstrate that wildfire resilience is context-dependent and requires region-specific restoration strategies rather than uniform interventions.

These archetypes are derived from a rule-based synthesis of dominant SHAP contribution rankings, providing a transparent and reproducible logic for classifying regional resilience. While currently based on qualitative synthesis due to the sample size, this framework links environmental drivers directly to successional outcomes.

4.5. Limitations and Future Directions

Several limitations should be acknowledged when interpreting these results. Primarily, the recovery indicator (ΔNDVI) primarily captures canopy greenness and cannot fully distinguish structural recovery from compositional changes. Post-fire ecosystems may exhibit rapid spectral greening dominated by herbaceous species while woody vegetation remains absent. Future studies should integrate structural remote sensing data (e.g., LiDAR or radar) to improve ecological interpretation. Additionally, while aggregated temperature and precipitation capture macro-environmental forcing, they overlook fine-scale soil moisture dynamics and unquantified human interventions (e.g., grazing and replanting). These factors likely explain the declining explanatory power of our models during later successional stages, highlighting the need to incorporate ecohydrological and management variables in the future.

Methodologically, the machine-learning attribution framework identifies statistical importance rather than direct causality. Although SHAP improves interpretability, causal relationships among climate, landscape structure, and biological processes require integration with process-based ecological models. Closely related to this methodological constraint, while the identified recovery archetypes provide mechanistic insights, their classification was based on an interpretative synthesis of SHAP results. Because the current study focuses on five case sites, utilizing formal quantitative clustering algorithms was neither necessary nor statistically robust. However, as the sample size expands in future research, objective clustering algorithms should be employed to enhance the robustness of these patterns. Additionally, the exclusion of variables like initial fire severity, pre-fire land-cover, and climatic history limits the model’s causal explanatory power. Consequently, the observed shift from climate to topography control should be framed as a conditional inference based on the available predictors, rather than an exhaustive description of recovery mechanisms.

Lastly, the temporal coverage of post-fire observations varies across sites. More recent fires, such as those in Yajiang, have shorter post-fire monitoring periods, so the full recovery trajectories are not yet observable. This limits direct comparison with older cases and may temporarily underestimate recovery duration, rate, and extent. While NDVI standardization allows relative comparison, interpreting spatial heterogeneity of resilience across fires must account for these temporal constraints. Future work should therefore extend monitoring of recent fires and incorporate methods to account for incomplete recovery trajectories to improve cross-site comparability.

In conclusion, addressing these limitations will allow future research to move from pattern detection toward predictive and mechanistic modeling of ecosystem resilience under increasing fire disturbances.

5. Conclusions

This study integrates high-resolution remote sensing, time-series vegetation dynamics, and explainable machine learning to advance the understanding of wildfire impacts on land systems. Accurate spatial identification using Sentinel-2 is shown to be critical for interpreting ecosystem recovery processes, particularly in fragmented and topographically complex landscapes where traditional datasets systematically overestimate or omit affected areas. Furthermore, post-fire vegetation trajectories demonstrate strong spatial heterogeneity. The results indicate that wildfire effects on land ecosystems are not solely driven by the disturbance itself but are heavily conditioned by the pre-existing environmental context. The explainable machine learning models reveal a general transition from early hydrothermal control to later topographic regulation, confirming that post-fire recovery is regulated by interacting nonlinear constraints rather than single drivers.

From a land management perspective, these findings emphasize that restoration strategies must be spatially differentiated. Climate-sensitive areas require early intervention to prevent degradation, high-altitude regions need long-term moisture conservation measures, and heterogeneous landscapes benefit from connectivity-oriented planning. Practically, this integrated approach offers a scalable operational tool for wildfire monitoring and land management. It enables local authorities to rapidly identify high-risk degradation zones shortly after a fire and optimize the allocation of limited restoration resources. By linking precise burned-area mapping with mechanistic recovery analysis, this study provides a transferable framework for post-fire assessment, ecological restoration prioritization, and adaptive land-use planning under increasing wildfire disturbance.