Unraveling Interactive Effects of Climate, Hydrology, and CO₂ on Ecological Drought with Interpretable Machine Learning

Abstract

As the risk of drought increases due to climate change, understanding ecological drought has become increasingly important for ensuring water resource security and carbon balance. However, most current ecological drought assessments rely on meteorological or hydrological indicators, which may not accurately reflect changes in the eco-physiological status of ecosystems. Therefore, this study establishes an ecological drought assessment framework using solar-induced chlorophyll fluorescence (SIF) as an indicator to examine its interpretable responses to climate–hydrology–environmental variables. The framework was tested across China’s nine major river basins and different ecosystems. Results show that SIF increased in 80.0% of China’s areas, with 60.9% showing significant increases (p < 0.05). Forest ecosystems experienced the lowest frequency of ecological drought but showed increasing duration and intensity, while grassland ecosystems had the highest frequency but decreasing duration and intensity. LightGBM machine learning analysis revealed that surface soil moisture (SMs), temperature (Tm), root-zone soil moisture (SMrz), and CO₂ were the main factors influencing ecological drought, with SMs and Tm contributing to over 66.1% of ecological drought. The SMs-Tm interaction alleviated ecological drought under low-temperature and high-humidity conditions but initially intensified then alleviated ecological drought under high-temperature and high-humidity conditions. The SMs-CO₂ interaction promoted ecological drought at high or low CO₂ concentrations but alleviated it at moderate concentrations.

1. Introduction

Since the 21st century, global climate change has intensified, leading to more frequent and severe extreme drought events [1,2,3]. These droughts typically last longer and show greater intensity, causing water shortages with cascading effects [4,5,6]. They exert multiple influences on meteorological, agricultural production, and hydrological cycles, thereby increasing pressure on terrestrial ecosystems [7,8]. Notably, driven by both climate change and human activities, the global vegetation greening trend has become increasingly prominent [9,10]. This change may further alter terrestrial ecosystems’ response to drought [4,11]. With growing challenges to terrestrial ecosystems and increasing societal emphasis on ecological protection, ecological drought has emerged as an important focus in drought research [12,13,14].

The characterization methods of ecological drought show a variety of characteristics [15,16,17]. The Sixth Assessment Report of the Intergovernmental Panel on Climate Change (IPCC) defines it as ecological feedback processes related to vegetation water stress. Reference [18] proposed that ecological drought is an occasional deficit in water resource availability that causes the ecosystem to exceed the vulnerability threshold. While expressions may vary, the academic community generally agrees that vegetation water stress is the core element characterizing ecological drought [19,20,21]. However, most studies currently use meteorological drought indicators (such as SPEI, VPD, etc.) to replace ecological drought, often ignoring the health dynamics of the ecosystem itself [13,15]. More importantly, gross primary productivity and the vegetation health index can reflect the health dynamics of vegetation and are often used to characterize ecological drought [22,23]. However, these vegetation indicators do not capture the changes in the eco-physiological functions of the vegetation ecosystem [24].

In recent years, research on ecological drought has expanded from focusing on the influence of single vegetation growth to multiple dimensions. This trend has been specifically manifested as follows: (1) hydrological dimension: such as the evaporation stress index based on the coupling relationship between actual evaporation and potential evaporation [24,25]; (2) ecological water cycle dimension: such as the ecological water loss index, which considers the water balance of vegetation supply and demand [13,26]; and (3) physiological mechanism dimension: such as the ecological drought index that reflects the physiological status of vegetation [27,28]. However, most of the existing indicators are comprehensive representations of vegetation’s response to drought, which poses challenges for the precise monitoring of ecological drought [28,29]. In the context of intensified climate change and the continuous degradation of ecosystem service functions, it has become an urgent task to construct a drought monitoring framework that can truly reflect the ecological and physiological functions of vegetation.

Solar-induced chlorophyll fluorescence (SIF) is the fluorescence signal emitted by chloroplasts in vegetation in the 650–800 nm band after absorbing photosynthetically active radiation under natural light conditions [30,31]. It is regarded as a natural probe for monitoring the photosynthetic activities of vegetation because it is directly related to the intensity of photosynthesis [32,33]. This technology has demonstrated significant application value in areas such as the dynamic assessment of vegetation productivity, diagnosis of water stress, identification of phenological changes, monitoring of carbon cycles, prediction of cropland yields, and drought early warning [34,35,36]. Compared with traditional indicators, such as gross primary productivity and the vegetation health index, SIF has unique advantages, including high time sensitivity, the direct indication of physiological processes, and early response to stress. It excels in monitoring vegetation water stress and evaluating eco-physiological functions [37,38]. However, current research on SIF mainly focuses on meteorological drought response and drought warning, with its application in ecological drought assessment remaining insufficient [36,39].

Therefore, this study aims to construct an innovative ecological drought assessment framework, which consists of three key components: first, analyze spatiotemporal dynamic characteristics based on SIF indicators; second, construct a new ecological drought index system using run theory to systematically evaluate the frequency, duration, and intensity of drought events; and finally, adopt interpretable machine learning methods to deeply analyze the driving mechanisms and interactions of multiple factors such as climate, hydrology, and environment on ecological drought. This framework will be applied and validated in China, a typical region significantly affected by climate change and vegetation greening. The research results can not only enhance the adaptability of terrestrial ecosystems to climate change and human activities but also provide a scientific basis and practical reference for ecological protection and sustainable development in other regions worldwide.

2. Materials and Methods

2.1. Study Area

China (The People’s Republic of China) has a land area of approximately 9.6 million km². From 2001 to 2022, the average precipitation was 636.7 mm, and the average temperature was 10.1 °C (Figure 1). During this period, temperature showed a significant upward trend (p < 0.05) (Figure 2d). The area proportions of cropland, forest, and grassland ecosystems in China are 19.9%, 25.4%, and 29.4%, respectively (Figure 1b). To address land degradation, air pollution, and climate change, China has implemented various ecological restoration projects, covering a cumulative area of 6.2 million km² with a total investment of USD 378 billion [40,41,42]. These measures have significantly increased vegetation coverage and led to extensive greening. The forest area in China continues to show a significant upward trend (p < 0.05) (Figure 2c). Furthermore, this study divides China into nine sub-regions: the Northwest Inland River Basin (NW), the Songliao River Basin (SL), the Yellow River Basin (YR), the Hai River Basin (HA), the Huai River Basin (HU), the Southwest River Basin (SW), the Yangtze River Basin (YZ), the Southeast River Basin (SR), and the Pearl River Basin (PR).

2.2. Data and Processing

2.2.1. Land Cover and SIF Dataset

The land cover dataset was obtained from Reference [43]. The time scale adopted was annual data from 2001 to 2022, with a spatial resolution of 30 m × 30 m. The nearest sampling method was used to re-interpolate the data to a 0.1° × 0.1° grid. The areas of cropland, forest, and grassland with unchanged coverage from 2001 to 2022 were calculated for subsequent analysis. The SIF dataset was obtained from Reference [44]. The time scale adopted was 4-day interval data from 2001 to 2022, with a spatial resolution of 0.05° × 0.05°. Bilinear interpolation was used to re-interpolate this data to a 0.1° × 0.1° grid for further analysis. The ecological restoration data came from Reference [40].

2.2.2. Climate and Hydrological Dataset

The data for precipitation (Pre), temperature (Tm), and relative humidity (Rhu) were obtained from the CN05.1 dataset, and the vapor pressure deficit (VPD) was calculated using the following formula.

$$\mathrm{VPD}=e_s-e_a$$

(1)

$$e_s=0.611\exp ({\displaystyle \frac{17.27\mathrm{Tm}}{\mathrm{Tm}+237.3}})$$

(2)

$$e_a={\displaystyle \frac{e_s\mathrm{Rhu}}{100}}$$

(3)

where $e_s$ is the saturated vapor pressure, $e_a$ is the actual vapor pressure, Tm is the temperature, and $\mathrm{Rhu}$ is the relative humidity.

The data for surface soil moisture (SMs) and root-zone soil moisture (SMrz) were obtained from the GLEAM4.2a dataset (https://www.gleam.eu, accessed on 9 October 2024). The time scale of these data was daily, covering the period from 2001 to 2022. The CN05.1 and GLEAM datasets were processed into 4-day time scale data by averaging over 4 days. CN05.1 was bilinearly interpolated from 0.25° × 0.25° to 0.1° × 0.1°.

2.2.3. CO₂ Dataset

The CO₂ dataset was derived from Jena CarboScope (https://www.bgc-jena.mpg.de/CarboScope/, accessed on 9 October 2024). The time scale of the dataset was daily, covering the period from 2001 to 2022. After averaging over 4 days, the data were processed into 4-day time scale data, and bilinear interpolation was performed to convert them to a 0.1° × 0.1°

2.3. Spatial Trend Analysis

Spatial trend analysis and significance testing were mainly conducted using Sen’s slope and the Mann–Kendall trend test [45,46]. In the Sen’s slope trend analysis method, the magnitude of β is the primary indicator for measuring trend changes.

$$\beta =Median({\displaystyle \frac{x_i-x_j}{i-j}})\forall j>i$$

(4)

where a β value greater than 0 indicates an increasing trend, while a β value less than 0 indicates a decreasing trend.

The steps for conducting the Mann–Kendall trend test are as follows.

$$\mathrm{sgn}(\theta )=\left\{\begin{array}{ll}1,& \theta >0\\ 0,& \theta =0\\ -1,& \theta <0\end{array}\right.$$

(5)

$$S={\displaystyle {\sum }_{i=1}^{n-1}{\displaystyle {\sum }_{j=i-1}^n\mathrm{sgn}(x_j-x_i)}}$$

(6)

where n is the sample size, sgn represents the sign function, and $x_j$ are the j-th and i-th values of the sample sequence, respectively. S is the statistic that follows a normal distribution. The standardized statistic of the time series is defined as Z.

$$Z=\left\{\begin{array}{cc}{\displaystyle \frac{S-1}{\sqrt{Var(S)}}},& S>0\\ 0,& S=0\\ {\displaystyle \frac{S+1}{\sqrt{Var(S)}}},& S<0\end{array}\right.$$

(7)

$$Var(S)={\displaystyle \frac{n(n-1)(2n+5)}{18}}$$

(8)

2.4. Characteristics of Ecological Drought

Considering that SIF reflects vegetation photosynthesis, we used SIF anomalies to characterize ecological drought. Unlike other studies [17,23], these studies conducted ecological drought research on a monthly scale. We employed 4-day scale data for ecological drought identification, which provided to be a more detailed reflection of its changing characteristics. Based on previous research [23], we applied Equation (9) to calculate the SIF anomaly.

$$Std{\mathrm{SIF}}_{(y,c)}={\displaystyle \frac{{\mathrm{SIF}}_{(y,c)}-mean({\mathrm{SIF}}_{(2001-2022,c)})}{sd({\mathrm{SIF}}_{(2001-2022,c)})}}$$

(9)

where y represents the years from 2021 to 2022, c refers to the 4-day time scale, and mean and sd represent the mean and standard deviation of the c-time scale SIF from 2021 to 2022, respectively.

Ecological drought is defined as the $Std{\mathrm{SIF}}_{(y,c)}$ having at least three 4-day periods where its value is less than 0. Ecological drought on a 4-day scale is susceptible to environmental changes, causing short-term fluctuations. Previous studies on ecological drought have primarily focused on monthly-scale analyses, which often overlook finer temporal dynamics. SIF, as a direct proxy for photosynthetic activity, can capture vegetation physiological responses at substantially higher temporal resolution. In this study, we define ecological drought occurrence when SIF values remain below the threshold for three consecutive periods (i.e., >12 days). This approach enables the detection of drought impacts at a more physiologically relevant time scale than conventional monthly assessments.

Referring to the pooling method for drought events, we combined two ecological drought events with a 4-day interval. Considering that the main vegetation activity occurs during the growing season (from April to October), subsequent ecological drought research primarily focused on changes in SIF during this period. In this study, we conducted a quantitative analysis of the three characteristics of ecological drought: frequency (the number of occurrences), duration (the length of time the drought persists, measured on a 4-day scale), and intensity. The intensity of ecological drought is calculated as the absolute value of the cumulative drought index divided by the number of drought events. It is unitless, and a higher value indicates a more severe ecological drought.

2.5. Interpretable Machine Learning (IML) Model

IML is an advanced method for elucidating the nonlinear relationship between individual predictions and model decisions, and it has been widely used to enhance the understanding of complex earth systems [47,48].

2.5.1. Light Gradient Boosting Machine (LightGBM)

LightGBM is designed for the efficient training of gradient boosting models on large datasets. Compared to other boosting models, it uses a histogram-based algorithm to speed up the training process and adopts a leaf-wise growth strategy with depth constraints. LightGBM is based on gradient-based one-side sampling (GOSS) and exclusive feature bundling (EFB) processes to aid in training. The GOSS algorithm excludes a large proportion of data cases with small gradients, and the remaining data instances are used solely to estimate information gain. Specifically, variance gain can be applied to split data instances, as shown in Equation (10):

$$\stackrel{\sim }{V_j}(d)={\displaystyle \frac{1}{n}}\left({\displaystyle \frac{{({\displaystyle {\sum }_{x_i\in A_l}{g}_i+\frac{1-a}{b}}{\displaystyle {\sum }_{x_i\in B_l}{g}_i})}^2}{n_l^j(d)}}+{\displaystyle \frac{{({\displaystyle {\sum }_{x_i\in A_r}{g}_i+\frac{1-a}{b}}{\displaystyle {\sum }_{x_i\in B_r}{g}_i})}^2}{n_r^j(d)}}\right)$$

(10)

where $n_l^j(d)$ and $n_r^j(d)$ are the number of left and right nodes, respectively; ${\displaystyle \frac{1-a}{b}}$ is the coefficient for gradient normalization; $A_l$ and $A_r$ are subsets of subset A; $g_i$ is order gradient statistical results of loss functions; $B_i$ and $B_r$ are subsets of subset B.

In addition, EFB can reduce the number of features by bundling mutually exclusive features in sparse feature spaces, thereby improving model performance and computational efficiency. For more detailed information, refer to Reference [49]. Furthermore, we implement this step in Python 3.10 using the LightGBM package integrated with scikit-learn.

2.5.2. Model Verification

We used 80% of the dataset for training and 20% for testing, and evaluated both the training and test sets using the Nash–Sutcliffe efficiency coefficient (NSE).

$$\mathrm{NSE}=1-{\displaystyle \frac{{{\displaystyle {\sum }_{i=1}^n(H_s-H_o)}}^2}{{{\displaystyle {\sum }_{i=1}^n(H_o-\overline{H_o})}}^2}}$$

(11)

where $H_o$, $H_s$, $\overline{H_s}$, and $\overline{H_o}$ are the observed, simulated, mean simulated, and observed variables, respectively; n is the number of data points.

The hyperparameter optimization of the LightGBM model primarily focused on the learning rate, n_estimators, and num_leaves. LightGBM machine learning model were employed to analyze the relationship between ecological drought and relevant climate–hydrological–environmental factors. The NSE values for the LightGBM model in the training and test datasets for all ecological drought events were 0.832 and 0.829, respectively. Specifically, for cropland-related ecological drought events, the NSE values were 0.797 and 0.778 in the training and test datasets. For forest-related ecological drought events, the NSE values were 0.869 and 0.864, while for grassland ecological drought events, the values were 0.764 and 0.705 (

Table 1. NSE of different ecosystems on training and test datasets.

Classification	China	Corp	Forest	Grass
Train	0.832	0.797	0.869	0.764
Test	0.829	0.778	0.864	0.705

). The LightGBM model performed the best in simulating forest ecological drought events and was least effective for grassland ecosystems. Nevertheless, the overall simulation results were deemed satisfactory and were subsequently utilized for further analysis.

2.5.3. SHAP Additive Explanations

We applied the advanced SHAP framework (TreeExplainer) to analyze the sensitivity of ecological drought to changes in climate, hydrology, and the environment [50]. TreeExplainer extends local interpretation (i.e., the influence of input features on a single prediction) to directly measure interaction effects using SHAP interaction values, which provide a richer form of local interpretation. Specifically, for a given sample x with M features, the SHAP addition mechanism decomposes a single prediction into feature contributions as follows:

$$f(x)=g(x^{\prime })={\phi }_0+{\displaystyle \sum _{i=1}^M{\phi }_i{x}_i^{\prime }}$$

(12)

where f and g represent the original model and the explanation model, respectively, and g should match the output of the original model for a simplified input x′ derived from x, which represents the mean value of the prediction values.

The SHAP method can also provide a global explanation. The average of the absolute SHAP values is considered the importance of a given variable and can be calculated as follows:

$$I_{\mathrm{SHAP}}^j={\displaystyle \frac{1}{n}}{\displaystyle \sum _{i=1}^n\left|{\phi }_j^i\right|}$$

(13)

where j and i are the input variable and data sample, respectively, and n is the number of samples.

3. Results

3.1. Spatiotemporal Analysis of SIF

Figure 3 shows the spatiotemporal variations in SIF during the growing season. The SIF values are mostly distributed between 0 and 1.8, with an average value of 0.79. Spatially, the SIF values in northwestern and northern China are relatively low, while those in the eastern and southern regions are relatively high (Figure 3a). Statistical results for different river basins revealed that the average SIF values for the NW, YR, and SW basins were relatively low (0.21, 0.67, and 0.70, respectively), whereas the SL, HA, HU, YZ, SE, and PR basins had higher averages (1.15, 0.98, 1.19, 1.07, 1.17, and 1.27) (Figure 3b). Approximately 80.0% of areas showed a positive SIF trend, with 60.9% exhibiting significant increases, primarily concentrated in central and eastern China (Figure 3c). Basin-scale trend analysis indicated that the NW, HU, and SW basins had relatively low trend averages, while the SL, YR, HA, YZ, SE, and PR basins displayed stronger trends (Figure 3d). Notably, SIF spatiotemporal variations demonstrate significant differences across ecosystem types (cropland, forest, and grassland), necessitating further detailed analysis.

For the statistical results in China, the average SIF of forest was the highest at 1.22, followed by cropland, with a value of 1.11, and grassland, with the smallest value of 0.44 (Figure 4). The trend value of cropland was the largest at 0.010, followed by forest at 0.008, and grassland at 0.006. According to the statistical results of different river basins, the magnitude distribution of the mean SIF value in cropland was as follows: SW > HU > SL > PR > YZ > HA > SE > YR > MW. The magnitude distribution of its trend value was NW > SL > YR > PR > SE > YZ > HA > SW> HU. The magnitude distribution of the mean SIF of forest was PR > SL > YZ > HU > SE > SW > YR > HA > NW, and the magnitude distribution of its trend value was YR > HU > PR> YZ > HA > SE > SL > SW > NW. The magnitude distribution of the mean SIF for grassland was PR > HU > HA > SL > YZ > YR > SW > NW > SE, and the magnitude distribution of its trend values was HA > SL > YR > YZ > NW > PR >HU > SW > SE. There was no clear regional pattern in the changes in SIF values for different ecosystems.

3.2. Evolution Characteristics of Ecological Drought

Figure 5 shows the spatiotemporal distribution characteristics of the frequency, duration, and intensity of ecological drought during the growing season in China from 2001 to 2022. The frequency of ecological droughts during the growing season in China ranged from 0 to 68 occurrences, with higher frequencies mainly occurring in the northern river basins. The average distribution of the frequency of ecological drought in various river basins in China was HU > YR > HA > SW > NW > SL > YZ > SE > PR, while the distribution of its trend values was NW > YR > YZ > HA > HU > SE > SW > PR > SL. This indicated that, compared to the river basins in southern China, the frequency of ecological droughts in the northern river basins was decreasing more rapidly.

The duration of ecological droughts during the growing season in China ranged from 0 to 785 (unit: 4 days). The distribution of the mean duration of ecological droughts across different river basins in China was SL > HU > HA > YR > YZ > SW > SE > PR > NW. The trends of ecological drought duration in the SE, HU, YZ, PR, and SW basins were increasing, while those in the HA, SL, YR, and NW basins were decreasing. This suggested that the duration of ecological droughts during the growing season in northern China was decreasing, while in southern China, it was increasing.

The intensity of ecological droughts during the growing season in China ranged from 0 to 19.6. The average intensity distribution of ecological droughts across different river basins in China was SL > HA > YZ > HU > SE > YR > PR > SW > NW. The trends in the intensity of ecological droughts in the SE, PR, YZ, HU, SW, and HA basins were increasing, while in the YR, NW, and SL basins, they were decreasing. This indicated that the intensity of ecological droughts during the growing season in northern China was decreasing, while in southern China, it was increasing.

Figure 6 shows the evolution characteristics of ecological drought in different ecosystems (cropland, forest, and grassland) in China. The average number of ecological droughts occurring in grassland was the highest at 33 occurrences, followed by cropland at 29, and the lowest in forest at 25. The frequency of droughts in all three ecosystems (cropland, forest, and grassland) decreased, with a consistent reduction trend across the ecosystems. The duration of drought in cropland, forest, and grassland ecosystems was 491, 432 and 404 (unit: 4 days), respectively. While the duration of ecological droughts in forest showed an increasing trend, it decreased in both cropland and grassland, with grassland showing a greater decrease than cropland. The average drought intensity in forest was the highest at 11.0, followed by cropland at 10.6, and the lowest in grassland at 7.9. The trend of drought intensity in cropland and forest ecosystems showed an increase, with a greater increase in forest than in cropland. In contrast, the drought intensity in grassland showed a decreasing trend.

3.3. Interpretable Driving Mechanism of Ecological Drought

Figure 7 illustrates contribution rates of ecological drought drivers across various ecosystems by using LightGBM and SHAP. A positive SHAP value indicates a gain in SIF from the baseline value produced by a specific variable, while a negative value indicates a loss. The findings indicated that SMs, Tm, SMrz, Rhu, and CO₂ contributed to 47.0%, 25.0%, 11.2%, 5.4%, and 5.0% of all ecological drought events, respectively. Specifically, SMs, Tm, SMrz, and CO₂ accounted for 48.1%, 18.0%, 13.9%, and 6.4% of cropland-related ecological drought events, while Tm, SMs, SMrz, and CO₂ explained 38.7%, 31.2%, 10.1%, and 5.4% of forest-related ecological drought events. For grassland ecological drought events, SMs, Tm, SMrz, and CO₂ contributed to 48.2%, 22.6%, 10.6%, and 5.5%, respectively. SMs and Tm emerged as the predominant factors influencing ecological drought, with SMrz and CO₂ also playing significant roles. The factors affecting cropland and grassland ecological drought events (SMs > Tm > SMrz > CO₂) exhibited consistency. Notably, in contrast to cropland and grassland ecological drought events, temperature emerged as the most influential factor in forest-related ecological drought (Tm > SMs > SMrz > CO₂).

Two primary factors, SMs and Tm, as well as two significant factors, SMrz and CO₂, were utilized to elucidate the impact of climate–hydrology–environment factors on ecological drought by using SHAP additive explanations. The impacts of SMs and Tm on ecological drought were found to be consistent. As SMs and Tm levels increased, ecological drought tended to ameliorate, initially at a slow pace and then accelerated. Notably, Tm mitigated ecological drought at a slower rate during the initial phase and more rapidly during the subsequent phase compared to SMs (Figure 8a,b). SMrz exhibited a threshold effect on ecological drought, alleviating ecological drought when SMrz was below 0.15 and exacerbating it when SMrz exceeded 0.15 (Figure 8c). Similarly, CO₂ demonstrated a threshold effect on ecological drought, alleviating it when SMrz was below 0.015 or above 0.035, and intensifying it when SMrz was between 0.015 and 0.035 (Figure 8d).

Furthermore, we explored the impact of the interaction between soil moisture and climate–environmental variables on ecological drought by using SHAP additive explanations (Figure 9). When the temperature was high, the SHAP interaction value between SMs and Tm initially increased slowly and then decreased rapidly, indicating that the interaction between SMs and Tm was first promoted and then alleviated ecological drought. The SHAP interaction value between SMrz and Tm fluctuated when SMrz was greater than 0.35. When the temperature was relatively low, the SHAP interaction value between SMs and Tm decreased slowly at first and then rapidly, indicating that the interaction between SMs and Tm alleviated ecological drought. The SHAP interaction value between SMrz and Tm increased slowly, indicating that the interaction between SMrz and Tm promoted ecological drought. The interaction between SMs and CO₂ was quite complex. When the CO₂ was high or low (CO₂ < 0.015 and CO₂ > 0.035), the SHAP interaction value between SMs and CO₂ increased slowly, indicating that this interaction promoted ecological drought. When the CO₂ was moderate (0.015 < CO₂ < 0.035), the SHAP interaction value between SMs and CO₂ decreased slowly, indicating that the interaction between SMs and CO₂ alleviated ecological drought. There was no significant change in the SHAP interaction between SMrz and CO₂.

4. Discussion

4.1. Analysis of SIF Changes

Our research results showed that 80.0% of the area had a positive trend value for SIF, with 60.9% of the area showing a significant increase (p < 0.05), which was similar to the findings of Reference [35]. The changes in SIF characteristics were primarily driven by the combined effects of climate change and human activities. China was one of countries most severely impacted by both factors. Reference [51] found that the contributions of climate change and human activities to changes in photosynthesis of Asian vegetation accounted for 43.6% and 56.4%, respectively. Other studies also indicated that temperature was the main driving factor influencing photosynthesis [35,51]. Additionally, China’s ecological restoration projects, such as the Three-North Shelter Forest Program and the Grain for Green Project, led to vegetation greening, which was an important factor contributing to the positive changes in SIF [9,41].

4.2. Rationality of the Evolution Characteristics of Ecological Drought

This study employed SIF and run theory to characterize ecological drought patterns across China, revealing distinct trends in frequency, duration, and intensity. Our results align with the existing literature [17,27] demonstrating a nationwide decline in drought frequency post-2010, attributable to increased precipitation. Notably, northern China exhibited more rapid reductions in growing-season drought frequency compared to southern regions. While climate change and anthropogenic activities have generally exacerbated drought duration and intensity globally [52,53], our analysis identified divergent regional patterns: northern China’s growing season showed reduced drought severity due to ecological restoration and enhanced rainfall, whereas southern river basins experienced intensified droughts associated with rising temperatures. Vegetation-type analyses revealed forests had the lowest drought occurrence but increasing severity per event, whereas grasslands—predominantly in northern China [27]—recorded the highest frequency but decreasing severity. This spatial dichotomy reflects northern grassland improvements from vegetation recovery and precipitation increases, contrasted with southern forests where expanded greening efforts enhance drought resistance yet prolong drought impacts once initiated [54].

4.3. Driving Mechanism of Ecological Drought

The attribution of ecological drought evolution involves complex multi-factor interactions. Existing studies have approached this challenge through diverse methodological frameworks: hybrid machine learning techniques have been employed to quantify drought propagation probabilities from meteorological to ecological systems [55], while standardized regression coefficient analyses identified temperature as the dominant driver of ecological drought occurrence [17]. Beyond climatic influences, the groundwater–ecological drought relationship has been investigated through spatiotemporal matching approaches, particularly in arid regions like northwestern China [56]. Although these studies have systematically characterized ecological drought responses to individual meteorological and hydrological factors, they have not comprehensively addressed the integrated effects of coupled climate–hydro–environmental variables or provided mechanistic explanations. Our study bridges this critical gap by applying interpretable machine learning, revealing four primary determinants of ecological drought across China: SMs, Tm, SMrz, and CO₂. However, the primary drivers of ecological drought differ significantly between forest ecosystems and cropland/grassland systems. In croplands, the two dominant factors accounted for 48.1% (SMs) and 25.0% (Tm) of drought variability, respectively. Grasslands showed similar patterns with 48.2% (SMs) and 22.6% (Tm). Forests exhibited reversed dominance where Tm (38.7%) surpassed SMs (31.2%), though marginally. This demonstrates SMs’ overwhelming control in cropland/grassland systems (contributions ≈ 2 × Tm), while forests require co-occurring Tm and SMs stress for drought onset—reflecting their stronger drought resistance, as evidenced by References [35,51,57,58].

The impact of droughts on vegetation is essentially manifested as the transition of water shortage from the meteorological to ecological stages. For example, Park et al. [14] found 46 drought events are successfully paired with 130 meteorological and 184 ecological drought events during 1982–2020 in northwestern China [55]. The transmission of meteorological drought to ecological drought has a certain probability. Furthermore, the propagation from meteorological to ecological drought involves complex nonlinear processes. As summarized in Wang et al. [23], four distinct propagation patterns exist, i.e., (1) multiple meteorological drought triggering a single ecological drought, (2) a single meteorological drought inducing multiple ecological drought, (3) one-to-one event correspondence, and (4) cascading multi-event interactions. This complexity makes ecological drought verification through meteorological data alone inherently challenging. Furthermore, unlike hydrological drought that can be monitored through streamflow or reservoir levels, large-scale ecological drought lack direct observational proxies [28]. Currently, SIF has emerged as the most robust indicator for ecological drought assessment due to its direct linkage with vegetation photosynthetic activity. While existing research predominantly focuses on identifying ecological drought drivers and deciphering meteorological-ecological drought propagation mechanisms, the core manifestation of ecological drought remains vegetation physiological stress—which is best quantified through photosynthetic indicators like SIF and GPP [23,27,28].

4.4. Uncertainties and Limitations

This study was conducted within the framework of the spatiotemporal changes in SIF, the evolution characteristics of ecological drought indicators, and the driving mechanisms of ecological drought, providing valuable insights for related research on ecological drought. However, there are also some limitations in this study. The time scale adopted in this study was 4 days. Although this provides a more detailed characterization of ecological drought, most SIF products and other remote sensing products currently use a common 8-day scale. In the future, it will be important to use more data at the same time scale to reduce the uncertainty caused by different remote sensing data products. While this study successfully identified the key driving factors of ecological drought, it did not explicitly analyze the propagation mechanisms from meteorological drought or soil moisture stress to ecological drought systems. These critical transmission processes warrant further investigation to fully understand drought cascade dynamics. Future research directions should incorporate tower-based SIF measurements to validate both the occurrence and temporal development of ecological drought events, thereby strengthening the observational basis for drought monitoring.

5. Conclusions

This study revealed the spatiotemporal variations in SIF in China from 2001 to 2022, characterized the changing characteristics of the frequency, duration, and intensity of ecological droughts in China, and quantified the relative contributions of climate, hydrology, and environmental variables to ecological drought. The interaction effects of the main climate–hydro–environmental variables on ecological drought were explored. The main conclusions are as follows: (1) The proportion of areas with a trend value greater than zero for SIF is 80.0%, with the proportion of areas showing a significant increase at 60.9% (p < 0.05), mainly distributed in the central and eastern parts of China. (2) From 2001 to 2022, the number of ecological droughts in forests was the lowest; however, both the duration and intensity of ecological droughts showed an increasing trend. Ecological drought occurs most frequently in grasslands, but both the duration and intensity of its droughts show a decreasing trend. (3) SMs, Tm, SMrz, and CO₂ are the main factors affecting ecological drought in different ecosystems. In particular, the contributions of SMs and Tm to the changes in ecological drought exceed 66.1%. (4) With the decrease in Tm and the increase in SMs, the interaction between SMs and Tm alleviates ecological drought. As Tm and SMs increase, the interaction between these variables first promotes and then alleviates ecological drought. (5) When CO₂ levels are either high or low, the increase in SMs promotes ecological drought through their interaction with CO₂. When CO₂ is at a moderate level, the interaction between SMs and CO₂ alleviates ecological drought with the increase in SMs.

Our framework provides insights into understanding the driving mechanisms of ecological drought in the context of changing climate, hydrology, and environmental factors. It also offers a scientific basis and a practical reference for ecological protection and sustainable development in other regions around the world.