Identifying Forest Drought Sensitivity Drivers in China Under Lagged and Accumulative Effects via XGBoost-SHAP

Abstract

Drought, a complex and frequent natural hazard in the context of global change, poses a major threat to key forest ecosystems in the carbon cycle. However, current research lacks a systematic and quantitative analysis of the multi-factor drivers of drought sensitivity based on lagged and accumulative effects. To address this gap, a drought sensitivity model was established by integrating both lagged and accumulative effects derived from long-term remote sensing datasets. To leverage both predictive power and interpretability, the XGBoost–SHAP framework was employed to model nonlinear associations and identify the threshold effects of driving factors. In addition, the Geodetector model was applied to examine spatially explicit interactions among multiple drivers, thereby uncovering the coupling effects that jointly shape forest drought sensitivity across China. The results reveal the following: (1) Drought had lagged and accumulative effects on 99.52% and 95.55% of forest GPP, with evergreen broadleaf forest showing the strongest effects and deciduous needleleaf forest the weakest. (2) Evergreen needleleaf forests have the highest proportion of extremely high drought sensitivity (16.94%), while deciduous needleleaf forests have the least (1.02%), and the drought sensitivity index declined in 67.12% of forests over decades. (3) Temperature and precipitation are the primary drivers of drought sensitivity, with clear threshold effects. Evergreen forests are mainly driven by climatic factors, while forest age is a key driver in deciduous needleleaf forests. (4) Interactive effects among multiple factors significantly amplify spatial variations in drought sensitivity, with water–heat coupling dominating in evergreen forests and structure–climate interactions prevailing in deciduous forests.

1. Introduction

Droughts, as prevalent extreme climatic events within the context of global change, manifest considerable spatial and temporal variability [1]. Both the extent and severity of droughts have continued to rise, currently affecting over 50% of China’s land area [2]. These alterations have engendered a spectrum of deleterious impacts on water resources, agricultural output, and natural ecosystems [3]. From the perspective of drought typology, droughts are typically classified into meteorological, agricultural, hydrological, socio-economic, and ecological droughts. Meteorological droughts are often considered as the primary source of other types of droughts, which can propagate into agricultural and ecological droughts, ultimately causing socio-economic losses [4]. China’s vast territory and complex topography, spanning the tropics to the cold temperate zone, form a globally unique gradient of forest types. As the dominant terrestrial ecosystems, forest ecosystems cover approximately 24.02% of China’s land surface, absorbing 50% of the total land-based primary productivity of carbon dioxide annually and accounting for over 80% of the carbon storage in the entire terrestrial vegetation [5]. This feature gives forests irreplaceable ecological values in terms of climate regulation, carbon sequestration and biodiversity conservation.

However, under the dual pressures of climate warming, forest ecosystems are facing the risk of structural declines [6]. Notably, climate-driven changes can alter regional climate systems through chemical feedback. When droughts persist, photosynthesis in plants is limited, stomata close, and carbon assimilation decreases. Meanwhile, respiration is often enhanced by high temperatures. These processes may cause forest ecosystems to shift from carbon sinks to carbon sources [7], which further increases atmospheric CO₂ levels. Higher CO₂ concentrations drive global warming, which enhances potential evapotranspiration, intensifies water deficits, and causes more extreme precipitation patterns. These changes lead to higher drought frequency and severity [8]. Forests are both responders and feedback agents in this process. They are also strongly coupled with crop activity and land systems. Forest degradation reduces soil water retention and weakens local climate regulation. This increases the drought sensitivity of nearby farmlands. On the other hand, agricultural expansion, landscape fragmentation, and slope land reclamation increase disturbance at forest edges and further reduce forest stability [9]. These interactions drive multi-scale responses of ecosystems under drought stress. Therefore, against the background of frequent droughts and rapid land system changes, identifying the resistance of forests to drought and the associated driving mechanisms from the source has become a key to understanding the propagation of risks and management strategies in climate–ecological–agricultural systems.

Previous studies showed that droughts significantly impact the carbon sink capacity of forests. For example, moisture stress inhibits photosynthesis and alters respiration, leading to tree mortality, reduced vegetation cover, and potential transitions between carbon sources and reservoirs [10]. In a sensitivity analysis of different tree species to moisture changes, trees growing in wetter areas were found to be most sensitive to drought. Predictions show that forests growing in wetter and hotter areas may be at higher ecological risk under future climate change scenarios [11]. It was also shown that in parts of the world, conifers can be less drought tolerant than broadleaf species [12]. Mechanisms underlying ecosystem responses to meteorological drought stress are complex, and the timing and intensity of droughts have different impacts on carbon cycling and ecosystem services [13]. Compared with other meteorological drought indices, the Standardized Precipitation Evapotranspiration Index (SPEI) integrates the effects of precipitation and potential evapotranspiration changes on droughts. It retains the sensitivity of the Palmer Drought Severity Index (PDSI) to temperature, but also benefits from the simplicity of the Standardized Precipitation Index (SPI) in its calculation [14]. For example, based on the SPEI and the Normalized Difference Vegetation Index (NDVI; based on the reflectance difference between NIR and RED bands), Wang et al. [15] demonstrated that meteorological drought is positively correlated with vegetation growth in most areas of China. Moreover, Zhao et al. [16] found lagged and accumulative effects between vegetation growth and drought in the SPEI and NDVI correlation study, indicating that vegetation growth is not mainly driven by current water conditions but by early drought events [17].

NDVI and other vegetation indices are primarily constructed on the basis of the reflectance characteristics of vegetation [18]. They effectively capture the greenness of vegetation but cannot directly reveal the actual ecosystem growth [19]. In comparison with vegetation indices such as NDVI, the Gross Primary Production (GPP) estimated from the Solar-induced Chlorophyll Fluorescence (SIF), provides a more intuitive reflection of the impact of droughts on vegetation productivity [20]. Numerous studies on drought impacts were conducted using experimental designs and empirical and process-based modeling [21]. However these analyses cannot effectively quantify the sensitivity of vegetation to droughts under the influence of lagged and accumulative effects, especially in the different forest types [22]. In addition, partial MK partial correlation analysis, regression analysis, and other statistical analysis methods are often used to analyze the internal driving forces of the impact of droughts on vegetation productivity [23]. However, these methods make it difficult to effectively consider the threshold effect and couple effects of different driving factors.

This study employs SPEI and GPP datasets to characterize lagged and accumulative effects of meteorological droughts on forest ecosystems in China. We developed a Drought Sensitivity Index (DSI) to quantify forest responses to droughts. Furthermore, we proposed a dual-driver analysis framework that combined XGBoost-SHAP interpretation and the Geographical Detector to systematically assess the relative contributions and threshold effects of meteorological, hydrological, and vegetation conditions, and structural factors on the DSI to reveal the spatially varying mechanisms of drought sensitivity under multi-factor interactions. These findings contribute to a deeper understanding of mechanisms underlying forest drought responses under climate change and offer scientific support for adaptive forest management and the accurate evaluation of forest carbon sink capacity in the context of China’s dual carbon goals.

2. Materials and Methods

2.1. Study Area

China is located in the eastern part of Asia and has a vast territory with a total area of about 9.6 million km². The country’s topography is characterized by a three-tiered terrain, which includes mountains, plateaus, plains, hills, and basins. China’s climate types are strongly influenced by various factors such as geographic latitude, geographical environment, and land–sea positions. The annual precipitation in the southeastern coastal regions exceeds 1500 mm, whereas the arid inland areas in the northwest receive less than 200 mm. Mean annual temperature decreases from over 20 °C in the tropical south to below 0 °C on the Qinghai–Tibet Plateau [24]. As shown in Figure 1, North China, Northeast China, and parts of Northwest China exhibit lower multi-year average SPEI values, indicating frequent drought events, while southern regions show relatively wetter conditions.

China is one of the five most forested countries in the world. There is a forest area of about 2.31 million km², with a forest stock of over 21.4 million m³, and a forest coverage rate of 24.02% [25]. Under the combined influence of topography and climate, forest vegetation in China shows a distinct regional pattern, forming a diverse and continuous ecosystem gradient [26]. In Northeast China, the Greater Khingan, Lesser Khingan, and Changbai Mountains are major natural forest areas. They include deciduous broadleaf forest (DBF), deciduous needleleaf forest (DNF), and mixed forest (MF; composed of broadleaf–needleleaf and deciduous–evergreen mixtures). These forests are distributed in temperate monsoon climates, dominated by deciduous species with high seasonal productivity and strong carbon uptake. In the southwest, the Hengduan Mountains are the core forest zone, featuring vertical vegetation zonation with evergreen broadleaf forest (EBF) and evergreen needleleaf forest (ENF). EBF is mainly found in humid low-elevation areas, with complex canopy structures and high evapotranspiration, while ENF occurs at higher elevations, showing strong cold and drought resistance. In the southeastern coastal and southern regions, tropical–subtropical monsoon EBFs dominate. These forests have high species richness and biological productivity, and play a vital role in national ecosystem function [27]. In addition, CHF was used in this study to represent overall forest conditions at the national scale.

2.2. Data

2.2.1. GPP Data

The GPP dataset employed in this study is derived from the NIRv-GPP dataset furnished by the National Tibetan Plateau Data Centre (http://data.tpdc.ac.cn, accessed on 12 April 2025) [28]. The close association between SIF and photosynthetic processes can effectively respond to changes in plant photosynthetic activity, making it a powerful means of monitoring GPP, while the new vegetation index (NIRv) is highly correlated with remotely sensed SIF products and can be used as an alternative to SIF. The NIRv-GPP dataset combines AVHRR remote sensing data with flux tower data from numerous global sites, providing high-resolution GPP information with a long time series. This monthly dataset spans from 1982 to 2018 with a spatial resolution of 0.05°.

2.2.2. SPEI Data

High-resolution SPEI datasets are pivotal for evaluating and quantifying the intensity, persistence, incidence, and spatial extent of meteorological droughts. The SPEI datasets are effective in representing deviations from the mean climatic level, and thus characterize meteorological drought conditions at a variety of time scales. The MSWEP_hPET data from the global high-resolution drought dataset (https://essd.copernicus.org, accessed on 12 April 2025) published by Gebrechorkos et al. [29] are well coupled to the large-scale SPEI at the Asian level and provide comprehensive data. Therefore, this study used this dataset to analyze drought conditions and their duration in forest ecosystems across China. The dataset provides a time scale of 1–48 months with a spatial resolution of 0.05°.

2.2.3. Driving Factors

Six variables were used to analyze the drivers of vegetation sensitivity to meteorological droughts. Meteorological variables, including the mean temperature (Tem) and precipitation (Pre) dataset (http://data.tpdc.ac.cn, accessed on 12 April 2025; 1 km, monthly), were provided by Peng [30]. Hydrological factors included actual evapotranspiration (Eva) and root-zone soil moisture (SM) from the GLEAM v4.2a dataset [31], with a spatial resolution of 0.1° and monthly temporal frequency (https://www.gleam.eu, accessed on 12 April 2025), which accounts for soil moisture stress and has been shown to outperform some reanalysis datasets [32]. Vegetation conditions were characterized using the GIMMS3+ NDVI dataset (https://daac.ornl.gov, accessed on 12 April 2025; 0.0833°, semi-monthly), which provides the longest continuous record of vegetation dynamics [33]. Forest structural attributes were represented by a forest age (Age) dataset (https://figshare.com, accessed on 12 April 2025; 30 m, annual) from Shang et al. [34], and forest coverage (FC) dataset (https://zenodo.org, accessed on 12 April 2025; 30 m, annual) from Cai et al. [35]. Anthropogenic activity was represented by the nighttime-light (Nlt) dataset (http://data.tpdc.ac.cn, accessed on 12 April 2025; 1 km, annual), which provides radiance-like brightness values using deep learning reconstruction [36]. All datasets were resampled to a 0.05° resolution.

2.2.4. Land Cover Data

Due to the large time span of the study and the need to use more fine-grained land cover delineation system data, two datasets were used in this study, derived from the GLASS-GLC land cover data product [37] and the GLC_FCS30D product [38]. The GLC_FCS30D dataset provides more detailed land cover classifications at a 30 m spatial resolution. To ensure data compatibility, all datasets were resampled to a uniform resolution of 0.05°. By extracting areas where forest vegetation types remained unchanged from 1982 to 2018 and overlaying them with 2018 forest type boundaries, five distinct forest types were delineated as study regions for comparative analysis.

2.3. Methodology

2.3.1. Quantification of Lagged and Accumulative Effects of Meteorological Droughts

Pearson correlation analysis was used to investigate the strength and direction of the relationship between SPEI and GPP in this study. Considering that vegetation productivity is influenced by both the vegetation growth and temperature, GPP data from May to September each year were selected for the correlation analysis. To investigate the lagged effect of drought on forest GPP in China, the SPEI data at a 1-month time scale were selected with the GPP data for each year’s growing season, and the GPP of each pixel for each month were paired with the SPEI from a month earlier to form a series collection (1 ≤ a ≤ 12). By correlation analysis, 12 correlation coefficients R_a were obtained for each pixel, and the maximum lagged correlation coefficient R_lag with the corresponding number of lagged months T_lag was used to characterize the intensity and time scale of the lagged effect for that pixel [22]. The calculation formula is as follows:

$$R_a=corr\left(GPP,SPE{I}_a\right),1\le a\le 12$$

(1)

$$R_{lag}=\max \left(R_a\right),1\le a\le 12$$

(2)

where GPP represents the time series of NIRv-GPP for each month of the growing season from 1982 to 2018, and SPEI_a is the SPEI time series data with a lag of a month.

To study the accumulative effect of drought on forests in China, the SPEI data with time scales ranging from 1 to 12 months and the GPP data for the growing season were selected. The GPP of each month was paired with the SPEI from m months earlier to form a series collection (1 ≤ m ≤ 12). For each pixel, correlation analysis was performed, and the maximum accumulative correlation coefficient R_acc with the corresponding SPEI time scale T_acc was used to characterize the accumulative effect strength and time scale for that pixel. The corresponding calculation formula is as follows:

$$R_m=corr\left(GPP,SPE{I}_m\right),1\le m\le 12$$

(3)

$$R_{acc}=max\left(R_m\right),1\le m\le 12$$

(4)

where SPEI_m is the time series data of SPEI on an accumulative m-month time scale and R_m is the correlation coefficient on an accumulative m-month scale.

2.3.2. Sensitivity of Vegetation Growth to Meteorological Drought

The lagged and accumulative effects of drought on vegetation growth occur simultaneously and overlap, suggesting that vegetation growth is affected not only by the current month’s drought but also by the lingering effects of drought from previous months [39]. In order to measure the sensitivity of vegetation growth to meteorological droughts, it is necessary to evaluate by the intensity and time scale of drought impacts on vegetation. The high sensitivity of vegetation to drought is reflected in stronger drought impacts (R_lag, R_acc) on vegetation and faster response times (T_lag, T_acc) [40]. Therefore, a composite index, Drought Sensitivity Index (DSI), was used to comprehensively assess vegetation sensitivity to meteorological droughts, and the formula is as follows:

$$DSI=\sqrt{{\displaystyle \frac{{R_{lag}}^2+{(1-N{T}_{lag})}^2+{R_{acc}}^2+{(1-N{T}_{acc})}^2}{4}}}$$

(5)

where DSI is the sensitivity of vegetation to drought, and NT_lag and NT_acc are T_lag and T_acc normalized by the min–max normalization method. To further analyze the sensitivity of different forest types to drought, DSI was classified into five levels using the natural breakpoint method: very low sensitivity, low sensitivity, medium sensitivity, high sensitivity, and very high sensitivity.

The Sen’s slope estimator combined with the Mann–Kendall significance test was employed to analyze the trend of DSI from 1982 to 2018. The detailed calculation procedure can be found in Gan et al. [41]. The slope β > 0 indicates an upward trend, and β < 0 indicates a downward trend; the significance levels are categorized as not significant when p > 0.05, moderately significant when 0.01 < p ≤ 0.05, and highly significant when 0 < p ≤ 0.01.

2.3.3. Driving Analysis

To quantitatively analyze the mechanisms underlying the sensitivity of forest productivity to meteorological droughts in China, we employed the extreme gradient boosting (XGBoost) modeling approach, integrated with SHapley Additive exPlanations (SHAP) for model interpretation. This combined framework offers advantages in handling high-dimensional nonlinear variables, capturing multivariate coupling relationships, and enhancing model interpretability. It allows for a detailed assessment of the marginal contributions and directional influences of environmental drivers on vegetation productivity [42].

A temporal moving window approach was applied to enhance the model’s responsiveness to long-term climate trends while maintaining sufficient sample sizes [43]. Specifically, we used 20-year moving windows (18 in total) to compute the drought sensitivity index (DSI), which served as the response variable. For each window, average values of Tem, Pre, Eva, NDVI, SM, Age, FC, and Nlt were used as input features in the nonlinear regression model. Model training was performed using ten-fold cross-validation. The performance of the proposed model was compared with random forest and support vector machine models, and accuracy was assessed by RMSE and R² metrics [42]. The general predictive structure of the model is expressed as follows:

$${\widehat{y}}_i={\displaystyle \sum _{k=1}^k{f}_k}(x_i),f_k\in F(i=1,2,\cdots ,n)$$

(6)

where ${\widehat{y}}_i$ represents the predicted value for sample i, x_i is the input feature vector, K denotes the total number of regression trees, f_k is the k-th regression tree, and F denotes the set of CART tree functions.

To enhance model interpretability, this study adopted SHAP values. SHAP is based on the Shapley value concept from cooperative game theory. It quantifies the marginal contribution of each feature to the model output. For any given sample, the prediction can be expressed as the sum of the model’s baseline prediction and the SHAP values of all input features:

$$y_i=y_{base}+f(X_{i1})+f(X_{i2})+\cdots +f(X_{ij})$$

(7)

where y_base denotes the mean predicted value across all samples, and f(X_ij) represents the SHAP value of the j-th feature for the i-th sample, reflecting its marginal contribution to the model output. When f(X_ij) > 0, it means the feature has a positive promoting effect on the response variable; otherwise, it has a suppressing effect.

To further explore the spatial heterogeneity and coupling mechanisms of driving factors influencing forest DSI, we conducted a complementary analysis of the XGBoost–SHAP results. The analysis used the factor detector, interaction detector, and ecological detector modules from the Geodetector framework [44]. Specifically, the factor detector was used to quantify the explanatory power of individual variables on the spatial distribution of DSI; the interaction detector assessed whether combinations of two variables produced enhanced or nonlinear enhancement effects; and the ecological detector tested for significant differences in DSI among different variable groupings. The explanatory power of each factor was measured by the q-statistic, with higher q-values indicating stronger influence. All independent variables were discretized using the natural breaks method.

3. Results

3.1. Lagged and Accumulative Effects

3.1.1. Lagged Effect of Drought on Forest GPP

In Figure 2a, 99.52% of the vegetation GPP in China showed a positive correlation with lagged SPEI, and 34.04% of the area passed the significance test (p ≤ 0.05). The maximum lagged correlation coefficients showed a decreasing trend from north to south, with higher R_lag in the southern Sichuan Basin. As shown in the spatial distribution of T_lag (Figure 2b), forests exhibiting a six-month lagged effect to drought covered the largest area—approximately 17.63% of the study region—mainly distributed in the southeastern hilly areas and the Daxing’an Mountains. The second largest area corresponded to a seven-month lag (10.71%), while forests with shorter lagged times are mainly located in southwestern Yunnan.

Among all forest types, the highest intensity of the lagged effect of droughts occurred on EBFs, with an average R_lag of 0.12, followed by MFs, while DNFs showed the weakest response (Figure 3a). In addition, the longest average T_lag was also on EBFs, with as high as 6.68 months, followed by the DNF type with 6.58 months, and the lowest was on ENFs. This indicates that the lagged effect of droughts has a stronger impact on broadleaf forests. Regarding the significance of lagged effects, DBFs ranked highest, followed by MFs, and the lowest was observed in DNFs.

In Figure 3b, for EBFs, most vegetation showed T_lag at a 6–7 month time scale, during which the average R_lag was also the highest. ENFs primarily exhibited a 1-month lagged effect, followed by 5 months, with the 5-month scale showing the highest R_lag. DBFs and DNFs showed a similar pattern to EBFs, with 6–7 months dominating both in area and in average R_lag. MFs, as a mixed type, presented a diversified lag pattern, with relatively high proportions at 1, 6, and 8 months.

3.1.2. Accumulative Effects of Drought on Forest GPP

As shown in Figure 4a, 95.55% of the accumulated SPEI in China’s forest was positively correlated with GPP, with 48.47% of the positively correlated areas passing the significance test. The spatial distribution of the R_acc was similar to that of the lagged effect, showing a gradual decrease from north to south. Regions with higher R_acc were mainly located in the southeast hills and around the Sichuan Basin. Among the accumulative effects of drought on GPP (Figure 4b), the area with a 1-month time scale had the largest accumulative effect, accounting for 40.25% of the positively correlated area. The area with an accumulated time scale of 9–12 months accounted for 33.75%, while the area with a 4-month time scale accumulative effect was relatively small.

In terms of accumulative effects, drought has the strongest average impacts on EBFs, with a mean R_acc of 0.153. DNFs showed the weakest response, with an average R_acc of only 0.041. For the T_acc, DNFs had the longest average time scale at 7 months, followed by DBFs at 6.11 months, while EBFs had the shortest duration at only 3.86 months. The significance pattern of drought’s accumulative effects was consistent with that of the lagged effect. EBFs showed the most significant response, followed by MFs, while DNFs showed the weakest.

Figure 5b indicated that the T_acc across forest types were predominantly concentrated at the 1-month time scale. EBFs showed the most significant response, with 58.3% of the area affected and the highest R_acc occurring at 1 month. In ENFs, accumulative effects were distributed across a wider range beyond the 1–2month scale, with the average R_acc peaking at 7 months. Needleleaf forest types exhibited similar hotspot patterns, with notable accumulative effects not only at shorter time scales but also between 9 and 12 months.

3.2. Sensitivity of Vegetation Growth to Droughts

As shown in Figure 6a, areas with high and extremely high sensitivity account for 26.04% and 8.71%, respectively, and are mainly located in southern China, including the southeastern hills, southwestern Yunnan, and the southern Sichuan Basin. In contrast, 25.29% and 13.13% of forest areas showed low and extremely low sensitivity to drought, mostly distributed in the Daxing’an Mountains, Changbai Mountains, and the Western Sichuan Basin.

Different forest types exhibited varying drought sensitivity patterns (Figure 6b). From EBFs to ENFs, and then to DBFs and DNFs, the proportion of areas with extremely low and low sensitivity increased, while the proportion with high and extremely high sensitivity gradually decreased. Among them, ENFs showed the highest proportion of extremely high drought sensitivity at 16.94%, while DNFs had the highest share of extremely low sensitivity at 17.43%, and the lowest proportion of extremely high sensitivity at only 1.02%.

With the shift of the temporal moving window, both the lagged and accumulative effects of drought showed a weakening trend (Figure 7). Under this background, 67.12% of forests exhibited a decreasing trend in drought sensitivity. Among them, 31.88% showed a significantly decreasing trend. These forests were mainly located in the Daxing’an Mountains, Changbai Mountains, and Yunnan Province. In EBFs, areas with low and extremely low drought sensitivity increased, reaching 43.61% by the 18th window, while areas with high and extremely high sensitivity still accounted for 33.82%. In ENFs, the area with low and extremely low sensitivity increased slightly year by year. In deciduous forests (DBFs and DNFs), the area with high and extremely high drought sensitivity decreased sharply—from 38.73% and 54.17% in the first window to 13.15% and 5.80% in the 18th window, respectively. The trend in MF forests was more similar to that of the evergreen types (EBFs and ENFs), with a slight annual decline in drought sensitivity.

3.3. Driving Factors of Vegetation Sensitivity

To evaluate the predictive performance and stability of different models, we compared XGBoost, random forest, and support vector machine models across different forest types using 10-fold cross-validation. As shown in Figure 8, XGBoost consistently outperformed the other two models, achieving R² values above 0.7 for all forest types, with the highest R² of 0.78 observed in ENFs. The RMSE values of XGBoost remained consistently low across all forest types, further confirming its strong predictive accuracy. The error bars in the figure represent the standard deviation across folds, indicating that XGBoost not only achieved higher accuracy but also exhibited better stability.

As shown in Figure 9, in China’s forest, Pre and Tem were the dominant drivers. Higher Pre was associated with positive SHAP values, indicating that abundant rainfall enhanced DSI, while lower values showed a suppressive effect in some samples. In EBFs, SM had a higher contribution, but its SHAP value declined as SM increased. The sensitivity of ENFs to drought was mainly driven by Pre and Tem, where higher values led to higher sensitivity. In MFs, high Tem played a dominant role in increasing DSI, while DBF was more sensitive to changes in Pre. In contrast, in DNFs, greater Age and higher Eva were associated with lower DSI, suggesting that older stands and high energy conditions improve drought resistance. Nighttime-light, which represents human activity intensity, shows no significant driving effect on drought sensitivity across all forest types. This indicates that daily human activities have a limited impact on forest drought resistance.

Figure 10 illustrates the threshold effects of drought sensitivity drivers across China’s forests. Tem contributed positively to DSI mainly within the 15–22 °C range, indicating that warming under favorable temperature conditions enhances drought sensitivity. Pre significantly reduced forest sensitivity to meteorological drought in the mid-to-low range. At high values (>110 mm), it shifted to a positive contribution. This pattern shows suppression at low values and increases at high values. SM showed a slightly positive contribution, suggesting that under moist conditions, the forest ecosystem was more responsive to droughts. However, when SM exceeded 0.39 m³/m³, it effectively reduced drought sensitivity, continuing the threshold trend observed for Pre. Age also presented a threshold effect: when the Age is below 40 years, its contribution was primarily positive; as Age increases, drought sensitivity declined. In contrast, the effects of NDVI and Eva were relatively stable. NDVI showed a weak positive trend, indicating that better vegetation conditions may enhance the forest’s response to drought. Eva consistently exhibited a negative influence on DSI, especially under high evaporation conditions. FC and Nlt exhibit no significant threshold effects and contribute little to the DSI.

The interaction detection of driving factors revealed that the dominant drivers were consistent with those identified by the XGBoost–SHAP analysis, validating the reliability of the driving force interpretation method. As shown in Figure 11, interactions among all factors across all forest types exhibited an enhancing effect, indicating that multi-factor synergy significantly intensified the influence on the spatial variability of DSI.

At the national scale, the three strongest interactions were Pre ∩ Tem (q = 0.215), SM ∩ Tem (q = 0.208), and Eva ∩ Tem (q = 0.207), all of which were linear enhancements. This indicates that the combined effects of water availability and thermal conditions are the dominant mechanism shaping the spatial distribution of drought sensitivity across China’s forests. Among forest types, DBF showed the most pronounced interaction effects. Multiple factor combinations had q values greater than 0.22. The strongest interactions included Pre ∩ Tem (q = 0.245) and Age ∩ Pre (q = 0.235). All interactions exhibited nonlinear enhancement. These results highlight the sensitivity of DBF to both moisture disturbances and structural conditions. Notably, in DNF, the strongest interactions were also nonlinear, such as Age ∩ Pre (q = 0.199) and Age ∩ Tem (q = 0.173), underscoring the critical role of stand age in enhancing resistance to meteorological drought.

4. Discussion

4.1. Response of Forest Productivity to Droughts

“Meteorological drought, which causes moisture deficits, adversely affects the growth of forest ecosystems in China.” It produces lagged and accumulative effects with varying intensities and time scales, reflecting the different sensitivities of forest types to droughts. In general, evergreen forests have a longer response time to drought compared to deciduous forests, while broadleaf forests (EBFs and DBFs) exhibit stronger accumulative effects than coniferous forests (ENFs and DNFs).

Lagged droughts often have more deleterious effects than the initial drought [45]. EBF has the strongest lagged effect, the highest average lagged correlation coefficient, and the longest lagged time. This finding aligns with the findings of Qiaoqian et al. [46], which may be attributed to the ecological adaptation characteristics of EBF. These forests generally thrive in humid environments with strong water retention capabilities, through their larger biomass regulation systems. EBFs can sustain photosynthesis over time, and short-term droughts do not significantly affect biomass in a short lag period [47]. However, due to their long-term adaptation to humid conditions, fluctuations in moisture levels have a relatively stronger impact on their growth. In contrast, DBFs exhibit shorter lagged response. Once drought exceeds its physiological tolerance range, water deficit rapidly leads to stomatal closure. Leaf abscission then rapidly reduces photosynthesis [48]. This process exacerbates productivity loss. The findings are similar to those of D’Orangeville et al. [17]. In coniferous forests, ENFs retain their leaves for continuous photosynthesis and transpiration throughout the year. They may preferentially use shallow soil moisture when adapting to drought conditions [49]. This makes them more responsive to soil moisture with shorter lagged times. The lagged time in DNFs was intermediate between ENFs and DBFs, suggesting that their short-term response is faster when drought occurs. However, since deciduous leaves also reduce water consumption to maintain internal water balance, the overall response of DNFs is slightly delayed compared to DBFs. Due to the most challenging growth environment among all forests, DNFs use a growth strategy that is more suitable for water-scarce environments. This strategy results in the weakest lagged intensity.

The accumulative effect reflects the ability of vegetation to tolerate prolonged droughts. The intensity of the accumulative effect in EBFs is stronger, while the average accumulated time scale is shorter, suggesting a higher sensitivity and greater physiological damage under sustained drought conditions, which aligns with the findings of Han et al. [50]. This may be attributed to the fact that the EBF is adapted to high-moisture environments, and when drought persists, its root system is unable to efficiently replenish water, leading to the long-term suppression of photosynthesis [51]. Similarly, the accumulative effect in the DBF, a broadleaf forest, is also stronger. Broadleaf forests require significantly more water recharge during the growing season compared to coniferous forests [52]. Prolonged droughts cause them to experience repeated water stress. This stress affects their carbon sink capacity. In contrast, coniferous forest species generally have a well-developed deep root system and leaves covered with a thick cuticle to reduce water evaporation [53]. They can efficiently utilize limited water resources to maintain basic physiological activities under prolonged water-limited conditions. As a result, they show a weaker accumulative effect. This may be due to the fact that the DBF type has a longer leaf longevity and higher energy input per unit leaf area, whereas the DNF has a shorter leaf longevity, which limits the strength of the accumulative effect in DNFs.

Because of the long-term growth environments, different forest types developed distinct growth strategies, which leads to varying lagged and accumulative effects, reflecting the different sensitivity of forests to drought. Different forest types exhibit distinct sensitivities to meteorological drought, reflecting long-term ecological adaptations and physiological traits [54]. The ENF shows the highest sensitivity. Although needle-like leaves reduce transpiration, the ENF maintains year-round photosynthesis, leading to high water demand [55]. Its weak hydraulic regulation further increases vulnerability under drought stress. The EBF ranks next in sensitivity. Adapted to humid environments, the EBF has high transpiration rates, shallow roots, and limited stomatal control [51], lacking effective drought resistance mechanisms. The DBF shows moderate sensitivity. It can shed leaves to reduce stress, but high water demand during the growing season and shallow roots make it sensitive to drought onset. The DNF is the least sensitive. It grows in dry or cold environments and adopts water-saving strategies such as short life cycles, small leaf area, and strong water regulation capacity [56]. These traits help DNFs maintain physiological function under water deficit, showing strong ecological adaptability. To summarize, forests in more humid growing environments are more sensitive to drought, and broadleaf forests exhibit greater sensitivity to drought than coniferous forests [17]. Mixed forests show an intermediate sensitivity to drought between broadleaf and coniferous forests, exhibiting stronger ecological adaptability, suggesting that the different complementary effects of different tree species may enhance their ability to resist drought [57]. This is a conclusion that aligns with the findings of Yu et al. [58].

4.2. Changes in Drought Sensitivity with Variations in SPEI

The drought sensitivity of forests in China shows an overall decreasing trend over the past decades, reflecting adaptive adjustments of forest ecosystems under long-term water stress [59]. The drought sensitivity of different forest types is jointly driven by multiple factors, including meteorological variables, vegetation conditions, and structural attributes. A comprehensive analysis of the driving factors indicates that Tem and Pre are the primary determinants, while SM, EVA, and NDVI play regulatory roles across forest types. In addition, Age exerts a significant influence, especially in deciduous forests.

The spatial heterogeneity of drought sensitivity across forest types fundamentally arises from differences in dominant drivers and complex response mechanisms [60]. At the national scale, Tem and Pre are the key controlling factors, reflecting a coupled water–energy regulation mechanism, which aligns with previous findings that climatic variables are the main drivers of vegetation responses to drought [61]. Tem within an optimal range intensifies DSI, suggesting that moderate warming exacerbates plant water deficit. In contrast, excessively high Tem induces stomatal closure to reduce water loss. Pre also exhibits nonlinear response characteristics. The severe deficiency of Pre has cultivated the resistance of forests to meteorological drought to a certain extent. When the Pre is suitable, forests become accustomed to a humid environment. Higher Pre may lead to trees changing their water harvesting strategies and relying more on surface soil moisture, resulting in increased sensitivity to drought [62]. SM directly affects water absorption at the root level and extends the influence of Pre. Under low to moderate moisture conditions, increased soil water enhances DSI. However, when the soil is sufficiently moist, meteorological droughts are unlikely to translate into soil drought, thus having little impact on forest ecosystems [63]. Eva shows a generally negative effect, especially in high-evaporation zones, indicating stronger drought resilience in vegetation adapted to high-energy environments. The NDVI exhibits a weakly positive trend, suggesting that better vegetation conditions may lead to heightened sensitivity to drought fluctuations, consistent with the findings of Ma and Yuan [43]. However, FC shows a weak driving effect. In high-density stands, increased competition for water may reduce drought resistance, while low-density stands lack effective shading and moisture retention. Overall, FC has limited influence on drought resilience. Similarly, Nlt, representing human activity intensity, shows no significant effect. This may be because human disturbance in forest areas is generally low, resulting in limited impact on forest drought response.

Dominant driving factors vary across forest types. In evergreen forests (EBFs and ENFs), drought sensitivity is primarily regulated by the interaction of Tem and moisture. For EBFs, SM plays a prominent role, with a high-moisture–low-sensitivity pattern, reflecting its strong dependence on water availability. Long-term stable photosynthesis requires stable water uptake as support [64]. The ENF shows stable responses across Tem and Pre gradients, indicating high climate adaptability. In deciduous forests (DBFs and DNFs), Pre is the dominant driver. The DBF is particularly sensitive to low and moderate rainfall, highlighting its high water demand during the growing season. This may be closely related to its ecological strategy of rapid leaf expansion and carbon assimilation during the growing season [65]. In DNFs, structural factors stand out, with Age and Eva showing negative effects on the DSI, implying that older forests and high-energy environments enhance drought resilience—likely due to deep root systems and stable energy allocation strategies. MFs are primarily temperature-driven. The DSI increases sharply under high-temperature conditions, suggesting that their multi-species composition may lead to increased resource competition and greater sensitivity to thermal stress [57].

Synergistic interactions among multiple drivers are crucial in regulating forest drought sensitivity. At the national scale, combinations of Tem with various water-related factors contribute the most, all exhibiting significant positive synergy. Su et al. [66] also pointed out that drought losses in China will double as temperatures rise. This suggests that compound effects among climatic factors are stronger than individual impacts. The joint fluctuations of water and heat amplify spatial patterns in drought sensitivity. These results reveal the strong responsiveness of forest ecosystems to compound stressors [60]. Interaction patterns differ across forest types. Evergreen forests exhibit weaker interactions, with drought sensitivity primarily modulated by combined climate factors. Under jointly elevated temperature and humidity, the DSI increases significantly, indicating that high water dependence and stomatal conductance led to stress responses under hot–humid conditions, as also concluded by Zhou et al. [67]. The MF shows no single dominant interaction. It exhibits multiple moderate interactions. This suggests a complex yet stable response strategy formed under multi-dimensional environmental gradients. The results support the notion of stronger ecological adaptability in mixed forests [68]. The growth environment of deciduous forests itself lacks water, and older trees usually have deeper and wider roots that can absorb water from deeper soil layers. Additionally, older trees are more mature and stable in physiological functions [69]. Therefore, in DBFs, drought sensitivity is driven jointly by temperature and moisture. Age ∩ Pre is the strongest interaction term. This indicates that tree age contributes significantly to the improvement of drought resistance in DBFs when disturbed by precipitation deficit. Especially in DNFs, Age ∩ Pre and Age ∩ Tem are the core interaction terms, and this “structure climate” interaction strengthens their overall resistance to drought. This reflects that older trees have greater drought resistance under dry and hot conditions.

The overall trend of forest drought sensitivity in China is on the decline, but this process exhibits significant spatial heterogeneity and differences under the interaction of multiple factors. For evergreen forests, their drought sensitivity is predominantly regulated by the coupling of climate factors. Strengthening meteorological monitoring and drought warnings, rationally allocating water resources, and optimizing forest layout and microclimate regulation mechanisms will help enhance their adaptability to water fluctuations under high-temperature arid composite environments. In contrast, deciduous forests exhibit outstanding structural factors under water stress, and aging trees can effectively enhance the system’s drought resistance under dry and hot conditions. Therefore, it is important to pay attention to the protection and management of middle-aged and elderly forest stands. By optimizing the age structure of forests, improving the development of the root system, and restoring the understory layer, its comprehensive resistance to precipitation deficit and high-temperature stress can be enhanced.

5. Conclusions

This study utilized the SPEI and GPP data to quantify the lagged and accumulative effects of droughts on forest productivity in China from 1982 to 2018. A dual analytical framework combining XGBoost-SHAP interpretation and the Geographical Detector model was employed to quantify the drivers of the drought sensitivity of forests. The research reveals distinct mechanisms by which short-term and prolonged droughts affect forest GPP, and demonstrates that the sensitivity measures derived from lagged and accumulative drought metrics more accurately capture the spatiotemporal persistence of drought impacts on forest productivity. Moreover, the integrated driver analysis not only identifies dominant influencing factors across forest types but also uncovers their nonlinear threshold responses and copula-based dependence structures. The results provide scientific support for forest carbon cycle research, drought risk assessment, and forest ecosystem management. The main conclusions were as follows:

(1)

Drought had significant lagged and accumulative effects on forest GPP, with clear differences across forest types. A total of 99.52% of forest GPP was positively correlated with lagged SPEI, predominantly with a 6–7 month delay. Similarly, 95.55% of regions showed a positive correlation between accumulated SPEI and GPP, with 40.25% responding within just 1 month. EBFs exhibited the strongest lagged and accumulative responses, while DNFs showed the weakest.

(2)

The drought sensitivity of China’s forests varied significantly across space, with high-sensitivity areas concentrated in the south. ENFs had the highest proportion of extremely sensitive areas (16.94%), while DNFs had the highest proportion of extremely insensitive areas (17.43%). Over the 37-year period, 67.12% of forested areas showed a declining trend in drought sensitivity.

(3)

Temperature and precipitation were the dominant drivers, both showing threshold effects. Specifically, temperatures between 15 and 22 °C and precipitation between 60 and 110 mm tended to increase sensitivity. Evergreen forests showed strong responses to the synergy of water and heat, deciduous forests were highly sensitive to low precipitation, and forest age played a dominant role in DNFs.

(4)

Multi-factor interactions amplified the spatial heterogeneity of drought sensitivity. At the national scale, the interaction between precipitation and temperature (q = 0.215) was dominant. Evergreen forests were primarily driven by climate coupling effects, whereas deciduous forests were jointly influenced by structural and climatic factors.

Despite the comprehensive analysis conducted in this study, several limitations remain. First, although Nlt was used as a proxy for human disturbance, this indicator mainly reflects light intensity in densely populated areas. It cannot fully capture non-residential human activities in forested regions. Some disturbances occur in remote forest interiors far from urban centers, which may lead to an underestimation of human impact. Second, both GPP and SPEI datasets used in this study are moderate-resolution remote sensing products. While appropriate for large-scale analysis, they may not capture fine-scale physiological responses in complex terrains or structurally diverse forests. Due to variations in species composition and traits, trees within the same forest type may respond differently to drought, yet such intra-type heterogeneity is difficult to reflect at the current spatial resolution. Moreover, the moderate spatial resolution may smooth localized drought signals, limiting the ability to fully capture forest responses to short-term or extreme drought events, which are often highly heterogeneous in space and time. Third, this study focused on long-term climate trends but did not fully account for extreme events, such as short-term heatwaves or severe droughts. These events can trigger rapid ecosystem changes or cause irreversible losses in forest structure and carbon storage. Future research should integrate higher-resolution ecological and climate data, along with more representative human activity indicators and policy-related information, to improve the understanding and attribution of forest drought responses.