Spatiotemporal Variations and Driving Factors of Evapotranspiration in Subtropical China from 2001 to 2020

Highlights

What are the main findings?

• Two 500 m, 8-day ET products with matching spatiotemporal resolution, MOD16 and PML-V2, were intercompared, and PML-V2 showed better agreement with ChinaFLUX observations for subtropical China.
• Annual ET in subtropical China increased significantly during 2001–2020, with clear south–north and coast–inland gradients; SWDown and LAI were the dominant controls, with northern subregions mainly energy-limited and southern subregions jointly regulated by vegetation and temperature.

What are the implications of the main findings?

• The dominant factors for ET changes can vary from south to north in subtropical China, suggesting the significance of weighing different variables in modelling ET and managing water resources in this region.
• Residual ET concentrated in urban and cropland areas may partly reflect anthropogenic influence, whereas in regions such as karst landscapes or complex terrain, it likely reflects unmodeled natural processes.

Abstract

Evapotranspiration (ET) is a key component of the terrestrial water and energy cycle, and its long-term dynamics are essential for regional hydrological assessment in subtropical China. In this study, two widely used satellite-based ET products, MOD16 and PML-V2, were selected for intercomparison because they provide consistent spatial (500 m) and temporal (8-day) resolutions. Validation against flux observations showed that PML-V2 performed better than MOD16 and was therefore used for subsequent analysis. Based on the 500 m, 8-day PML-V2 dataset, the spatiotemporal variation in ET in subtropical China during 2001–2020 was examined using the Theil–Sen slope estimator, Mann–Kendall test, and Hurst exponent. To identify the most relevant controls on ET variation, eXtreme Gradient Boosting (XGBoost) and SHapley Additive exPlanations (SHAP) were used to screen environmental factors and rank their relative importance. Multiple linear regression (MLR) was then applied only to the selected dominant factors to quantify their contributions. Residual analysis was used to distinguish climate–vegetation effects from residual influences, which could arise from human activities and unmodeled natural processes. The results showed that annual ET averaged 669 mm and increased significantly at a rate of 2.03 mm yr⁻¹ from 2001 to 2020, with an accelerated increase after 2010. Spatially, ET exhibited clear gradients from south to north and from coastal to inland regions. Downward shortwave radiation (SWDown) and leaf area index (LAI) were the dominant drivers over most of the study area, although their controls varied geographically, with northern subregions being more energy-limited and southern subregions being jointly influenced by vegetation and temperature. Residual ET trends largely coincide with cropland and urbanising areas, indicating a partial influence of human activities, while in subregions such as XM, complex terrain and hydrological heterogeneity suggest that unmodeled natural processes may dominate. These findings enhance understanding of ET dynamics in subtropical China and demonstrate the value of high-resolution remote sensing products for regional hydrological monitoring and driver attribution.

1. Introduction

Terrestrial evapotranspiration (ET) refers to the total flux of water vapour released from the land surface and vegetation to the atmosphere through evaporation and transpiration processes [1,2]. This process involves the phase change of water from liquid (or solid, e.g., ice) to vapour [3] and comprises three main components: evaporation from soil and open water, plant transpiration, and the evaporation of precipitation intercepted by vegetation canopies [4]. Globally, approximately 60% of terrestrial precipitation returns to the atmosphere via ET, and changes in ET can profoundly influence basin-scale water availability and ecosystem health [5]. ET is not only a key component of the climate system, but also an essential link connecting the water, energy and carbon cycles [6,7]. However, most hydrological studies to date have focused on the supply side of water resources (e.g., precipitation, snow, soil moisture and groundwater), while paying comparatively less attention to the demand side, namely vapour losses to the atmosphere through ET [8]. As a major pathway of land-surface water loss, accurate and reliable ET estimates are therefore critical for water resource management [9]. In the context of increasingly frequent extreme events such as floods, droughts and heatwaves, and growing mismatches between water supply and demand, it has become crucial to quantify water losses through ET associated with vegetation stress alleviation and to understand both sides of the water supply–demand balance [10,11].

Ongoing global warming is intensifying the hydrological cycle [12], with an increasing fraction of precipitation being partitioned into evapotranspiration rather than runoff [13]. Global ET products indicate that many regions worldwide have experienced significant changes in ET [14]. However, the impact of warming on ET does not necessarily manifest as a persistent, monotonic increase; both the direction and magnitude of ET changes remain highly uncertain across climate zones, seasons and land-surface conditions [15,16]. In addition, against the backdrop of global greening over recent decades, changes in vegetation structure have generally enhanced canopy interception evaporation and transpiration potential, thereby exerting sustained impacts on land–atmosphere water and energy exchanges [17]. Yet attributing ET changes solely to greening may overestimate long-term trends [18]. Meanwhile, the Intergovernmental Panel on Climate Change (IPCC) has highlighted that human activities have substantially altered the global water cycle since the mid-20th century [19]. Water-management practices such as irrigation not only modify basin water balances but may also influence near-surface climate and regional moisture recycling [20,21].

In recent years, remote sensing has been widely used to investigate the spatiotemporal variability of ET and its responses to driving factors at both global and regional scales [22]. However, ET estimation remains subject to inherent uncertainty, largely because satellite sensors cannot directly observe ET, making it difficult to parameterize the complex, intertwined physical and biological processes governing ET [23], and because point-scale ET observations are spatially sparse and limited in coverage [24,25]. These limitations may lead to substantial discrepancies among ET products in both trend direction and magnitude [1,26]. Among the available ET products, the MOD16 series has long been extensively applied in China because of its early release, long-term continuity and abundant applications, and has become an important data basis for studies of ET variability and attribution [9,18]. Meanwhile, PML-V2 replaces global reanalysis forcing with a meteorological forcing dataset tailored for China and explicitly incorporates water–carbon coupling within its modelling framework. It further applies dynamic temporal smoothing to LAI and simulates transpiration (Ec) and canopy interception evaporation (Ei) at the component level, thereby providing a more realistic representation of ecohydrological processes [7]. They represent two widely used and methodologically distinct medium-resolution remote-sensing ET datasets with identical temporal (8 d) and spatial (500 m) resolutions. Such medium-resolution products provide a useful balance between spatial detail and regional-scale applicability, making them well suited for investigating ET variability across heterogeneous landscapes. Meanwhile, their consistency in temporal and spatial resolution helps reduce scale-induced differences and allows the comparison to focus more directly on uncertainties arising from algorithm structure, forcing data and process parameterization. For these reasons, MOD16 and PML-V2 were selected in this study to examine ET trends and variations in subtropical China.

Approaches such as geographically weighted regression (GWR) [27] and the Geodetector model [28] have been widely used to quantify the combined effects of climate change and human activities on ET. However, these methods typically rely on multi-year mean values as inputs and thus cannot account for interannual trends in the driving factors, which represents an important limitation [29]. In contrast, multiple regression models can incorporate long time series of ET together with meteorological and vegetation variables, offering advantages for analysing ET drivers over extended periods. Nevertheless, few studies have comprehensively quantified long-term ET dynamics in subtropical China and its internal subregions, leaving the impacts of climate change and human activities on water resources in this region insufficiently understood [30]. Therefore, it is of clear practical significance to investigate how climate change and human activities influence ET across subtropical China and its subregions, to disentangle the interactive effects among environmental factors, and to elucidate the coupled mechanisms through which these drivers regulate ET.

To address these knowledge gaps, this study uses remote-sensing ET products at 500 m spatial resolution and 8-day temporal resolution for 2001–2020. We first evaluate their applicability in subtropical China and select the best-performing dataset, and then systematically characterise the spatiotemporal patterns and trends of ET across the subtropical region and its subregions. For driver identification and attribution, we employ an XGBoost–SHAP framework to screen key meteorological and vegetation variables and to diagnose potential nonlinear sensitivities. We further combine multiple linear regression with residual-based significance testing to quantitatively partition the contributions of climate and vegetation to ET changes, thereby providing a basis for elucidating the mechanisms underlying regional ET evolution.

2. Materials and Methods

2.1. Study Area

This study focuses on subtropical China, with the regional boundary defined according to the Data Center for Resources and Environmental Sciences, Chinese Academy of Sciences, spanning 21.33–33.91°N and 91.39–122.49°E (Figure 1). Within this extent, the subtropical zone is further subdivided into ten subregions: the middle–lower Yangtze River reaches (CJ), the upper–middle Han River basin (HJ), the Sichuan Basin (SC), the Guizhou Plateau (GZ), the Yunnan Plateau (YN), the Jiangnan Hills (JN), the Fujian–Guangdong Hills (MY), the Taipei region (TB), the southern Yunnan mountains and plains (DN), and the southern flank of the Himalayas (XM). These subregions together cover a broad area of southeastern China, including the middle–lower Yangtze River Plain, extensive hilly terrain in the southeast, and parts of the Yunnan–Guizhou Plateau. Topography is highly heterogeneous, with elevation increasing from east to west and forming a pronounced terrain gradient (Figure 1).

The region hosts diverse vegetation types and is one of the areas with relatively high forest cover in China. At the same time, subtropical China is densely populated and highly developed economically, with rapid urbanisation and strong demands for agricultural irrigation and industrial water use, which substantially affect the regional water supply–demand balance and ecosystem services. The climate is characterised by a typical subtropical monsoon regime. A systematic investigation of ET characteristics and their drivers in subtropical China is not only crucial for improving understanding of regional hydrological processes, but also provides a scientific basis for rational water resource regulation, adaptation to extreme climate events and the sustainable management of ecosystem functions.

2.2. Data Sources

Two ET datasets used in this study were derived from the PML-V2 0.1.8 and MOD16A2GF.061 products, both with a spatial resolution of 500 m and a temporal resolution of 8 days. Annual ET fields for each product were generated on the Google Earth Engine (GEE) platform by compositing the 8-day data to annual means. Meteorological and vegetation variables were used as explanatory drivers to investigate the controls on ET variability. Specifically, these variables were used for correlation analysis with ET, XGBoost–SHAP-based predictor screening, and subsequent multiple linear regression attribution. Precipitation (Precip) and monthly mean air temperature (Tair) datasets (1 km, 1901–2021) were obtained from the National Tibetan Plateau Data Center [31]. Specific humidity (Qair) was taken from GLDAS2.1 [32], which has a spatial resolution of 0.25° and a temporal resolution of 3 h. In addition, downward surface shortwave radiation (SWDown), vapour pressure (VAP), and wind speed (Wind_sp) were obtained from the TerraClimate dataset [33], with a spatial resolution of 1/24°. We also used volumetric soil water content (Soil_w) from the ERA5 reanalysis produced with the Integrated Forecasting System (IFS) of the European Centre for Medium-Range Weather Forecasts (ECMWF), spanning 1950 to the present at 0.1° spatial resolution [34]. Leaf area index (LAI) was obtained from the MOD15A2 product, with a spatial resolution of 500 m and an 8-day temporal resolution. All variables were restricted to the period 2001–2020 and resampled to 1 km to ensure a consistent spatial grid for regression.

Furthermore, ET estimates were validated against eddy-covariance observations from three ChinaFLUX sites [35], Dinghushan (DHS), Qianyanzhou (QYZ) and Xishuangbanna (XSBN), for the period 2003–2010. All flux data had undergone ChinaFLUX quality control procedures prior to use. Among these, DHS and QYZ are located within the study region, while XSBN lies at its margin. XSBN represents a tropical seasonal rainforest site, DHS a south subtropical monsoon mixed forest site, and QYZ a mid-subtropical evergreen coniferous plantation site.

2.3. Methodology

In this research, Theil–Sen median trend analysis, the Mann–Kendall significance test and the Hurst exponent were used to characterise the spatiotemporal trends and persistence of ET in subtropical China. To investigate the effects of meteorological and vegetation factors and to separate the residual component potentially associated with human activities, we adopted a regression-based attribution framework combined with residual analysis. In this framework, an extreme gradient boosting (XGBoost) model with SHapley Additive exPlanations (SHAP) was applied as a diagnostic and screening tool to identify the dominant predictors and potential nonlinear sensitivities, thereby reducing subjective variable selection prior to multiple linear regression. Each method is described in detail in the following sections.

2.3.1. Theil–Sen Median Trend Analysis and Mann–Kendall Significance Test

The Theil–Sen median trend analysis and Mann–Kendall test can be combined to effectively judge the trend of long time series data and consequently have been extensively applied to analyse long time series data. The Theil–Sen estimator is a robust non-parametric method for estimating trends in time series [36,37]. This method can be used to compute efficiently and is insensitive to measurement errors and outliers. Therefore, it is applicable to trend analysis of long time series, with the following equation:

$$S_{ET}=Median\left({\displaystyle \frac{S_j-S_i}{j-i}}\right),\forall i<j$$

(1)

When $S_{ET}$ > 0, it reflects an increasing $ET$ trend; otherwise, it reflects a decreasing $ET$ trend.

The Mann–Kendall test is a nonparametric statistical method that ascertains the significance of a trend [38]. This method can determine whether a process represents a natural fluctuation or definite trend and is widely used for testing trends in hydrometeorological time series [39]. The calculation formula is as follows:

$$Z=\left\{\begin{array}{l}{\displaystyle \frac{S-1}{\sqrt{s\left(S\right)}}},S>0\\ 0, S=0,\\ {\displaystyle \frac{S+1}{\sqrt{s\left(S\right)}}},S<0\end{array}\right.S={\displaystyle \sum _{j=1}^{n-1}}{\sum }_{i=j+1}^{n}sgn\left(x_j-x_i\right)$$

(2)

$$sgn\left(x_j-x_i\right)=\left\{\begin{array}{c} 1, x_j>x_i\\ 0, x_j=x_i\\ -1, x_j<x_i\end{array}\right., s\left(S\right)={\displaystyle \frac{n\times \left(n-1\right)\times \left(2n+5\right)}{18}}$$

(3)

where $Z$ is the standardised test statistic and $S$ is the Mann–Kendall test statistic. Significance is evaluated at levels of 0.05 and 0.01. Based on both $S_{ET}$ and $Z$, $ET$ trends are classified into five categories: significant increase, marked increase, nonsignificant changes, marked decrease and significant decrease.

2.3.2. Hurst Index and R/S Analysis

The Hurst exponent, derived from rescaled range (R/S) analysis, is an effective metric for quantitatively characterising the long-term dependence of a time series [40]. The mathematical principle is as follows: for a time series $\left\{\xi \left(t\right)\right\},t=1,2,\dots n$, for any integer $\tau =1$, its mean sequence is defined as follows [41]:

$${〈\xi 〉}_{\tau }={\displaystyle \frac{1}{\tau }}{\displaystyle \sum _{t=1}^{\tau }}\xi \left(t\right) \tau =1,2,\dots n$$

(4)

$$X\left(t,\tau \right)={\displaystyle \sum _{\mu =1}^t}\left(\xi \left(\mu \right)-{〈\xi 〉}_{\tau }\right) 1\le t\le \tau$$

(5)

$$R\left(\tau \right)=\underset{1\le t\le \tau }{\underbrace{max}}X\left(t,\tau \right)-\underset{1\le t\le \tau }{\underbrace{min}}X\left(t,\tau \right) \tau =1,2,\dots n$$

(6)

$$S\left(\tau \right)={\left[{\displaystyle \frac{1}{\tau }}{\displaystyle \sum _{t=1}^{\tau }}{\left(\xi \left(t\right)-{〈\xi 〉}_{\tau }\right)}^2\right]}^{{\displaystyle \frac{1}{2}}} \tau =1,2,\dots$$

(7)

The relationship $R/S\propto {\tau }^H$ indicates that the Hurst phenomenon is present in the analysed time series, where $H$ is the Hurst index value. $H$ can be fitted by least squares regression according to $\left(ln\tau , lnR/S\right)$. Generally, $H$ refers to a constant ranging between 0 and 1. Time series that display a random process have $H$ close to 0.5. When $H$ is >0.5, the future trend is persistent with the prior period (i.e., continues to rise or decline), and the higher the $H$, the greater the persistence. When $H$ is <0.5, the future trend is likely to be anti-persistence in the next period (i.e., changing from rising to declining or vice versa), and the lower the $H$ is, the greater the anti-persistence is.

2.3.3. XGBoost–SHAP Framework for Identifying ET Drivers

XGBoost is an efficient machine learning algorithm based on gradient-boosted decision trees. Owing to its strong predictive skill, efficient parallel computing capability and robustness to missing data, it has been widely applied to analyse multivariate driving mechanisms in climatic, hydrological and ecological processes [42,43]. However, as a typical “black-box” model, XGBoost provides limited interpretability regarding its outputs and the relative importance of individual features. To address this limitation, we employ SHAP, a game-theory-based framework for model interpretation [44]. In this study, XGBoost–SHAP is employed as a diagnostic screening step to capture potential nonlinearities and interactions between climatic–vegetation variables and ET, and to rank predictors by their explanatory importance.

2.3.4. Multiple Linear Regression and Residual Analysis for ET Attribution

Multiple linear regression (MLR) is used to model the relationship between several independent (predictor) variables and a single dependent (response) variable, and to evaluate the extent to which each predictor explains the variance in the response. To better retain the characteristics of complex climatic and vegetation conditions, we first used the XGBoost–SHAP framework to identify the most important climatic and vegetation drivers, and then applied MLR to assess the influence of these dominant climatic and vegetation variables on ET. Based on the regression slopes, the main drivers of ET change were identified using a maximum-slope criterion. It is worth noting that, prior to applying MLR, all heterogeneous variables were standardised to eliminate the effects of differing units and scales and to ensure comparability among predictors. The main analytical steps are as follows:

First, an MLR model was constructed between observed ET (${ET}_{obs}$) and the selected climatic and vegetation variables to obtain a time series of predicted ET (${ET}_{pre}$) that reflects only the influence of climate and vegetation. Second, the residual ET time series (${ET}_{res}$) was then calculated as the difference between ${ET}_{obs}$ and ${ET}_{pre}$. This residual series represents the portion of ET variability not explained by the climatic and vegetation predictors. Then a significance test was applied to the ${ET}_{res}$ time series. Third, if the trend in ${ET}_{res}$ is not significant, this indicates that climatic or vegetation factors are the main controls on ET. Conversely, a significant trend in ${ET}_{res}$ suggests the presence of additional influences (e.g., noise or other factors, such as human activities) that exert a stronger effect on ET than the included climatic and vegetation variables.

$${ET}_{pre}=b_0+b_1{X}_1+b_2{X}_{2+}{b}_3{X}_3+b_4{X}_4$$

(8)

$${ET}_{res}={ET}_{obs}-{ET}_{pre}$$

(9)

where ${ET}_{pre}$ denotes the standardised predicted ET; $X_1$, $X_2$, $X_3$, and $X_4$ are the selected independent variables; $b_0$ is the intercept; and $b_1, b_2$, $b_3$ and $b_4$ are the standardised regression coefficients. Here, ${ET}_{pre}$ represents the climate-driven component of ET, ${ET}_{res}$ represents the residual component of ET, which may reflect contributions from human activities as well as other non-climatic factors that are not explicitly modelled, and ${ET}_{obs}$ denotes the observed ET.

3. Results

3.1. Spatiotemporal Characteristics and Future Trends of ET

3.1.1. Evaluation of Flux-Tower Observations and Gridded ET Datasets

The comparison of MOD16 and PML-V2 at an 8-day temporal resolution shows that the two datasets are highly correlated (r = 0.95) and exhibit broadly similar temporal dynamics (Figure 2a). Nevertheless, MOD16 shows more pronounced and stable seasonal peaks, whereas PML-V2 displays larger variability in peak magnitudes. In addition, almost every 8-day MOD16 value exceeds the corresponding PML-V2 value, resulting in significantly higher annual ET totals in MOD16 than in PML-V2.

To further evaluate the datasets, we used eddy-covariance ET observations from the XSBN, DHS and QYZ flux towers during 2003–2010 to validate the gridded MOD16 and PML-V2 ET products (Figure 2b). In terms of intra-annual seasonal differences, all three sites show a broadly consistent pattern, with the highest ET in summer and the lowest in winter. The contrast between summer ET and the other seasons is smallest at XSBN, intermediate at DHS and most pronounced at QYZ. Regarding interannual variability by season, winter ET exhibits relatively compact value ranges at all three sites, indicating limited year-to-year variability. At DHS and XSBN, interannual variability in spring, summer and autumn ET is also relatively modest, whereas at QYZ the distributions in these three seasons are highly dispersed, reflecting much stronger interannual fluctuations.

Overall, both products reproduce the general temporal patterns of ET at the three sites, with better performance at DHS and QYZ and the poorest agreement at XSBN. Nevertheless, PML-V2 consistently outperforms MOD16 at all three towers as demonstrated by the higher R² values and lower RMSE values at all three sites (Figure 2b). We therefore selected PML-V2 as the primary ET dataset for subsequent analyses in this study.

3.1.2. Spatiotemporal Patterns and Future Trends of ET

From 2001 to 2020, the annual total ET in subtropical China ranged from 638 to 703 mm, with a mean of 669 mm, and exhibited a significant increasing trend of 2.03 mm yr⁻¹ (p < 0.01) (Figure 3a). A turning point occurred around 2010, with ET showing only a slight increase during 2000–2009 (0.35 mm yr⁻¹), whereas the rate of increase became much stronger after 2010 (1.88 mm yr⁻¹). Spatially, the trends in annual ET differ markedly among the subtropical subregions (Figure 3c). Overall, most areas display a significant increasing trend (p < 0.05), in the order TB (12.75 mm yr⁻¹) > XM (7.58 mm yr⁻¹) > DN (6.11 mm yr⁻¹) > YN (5.40 mm yr⁻¹) > JN (2.25 mm yr⁻¹) > MY (2.12 mm yr⁻¹) > GZ (0.04 mm yr⁻¹). In contrast, several regions show decreasing trends, namely HJ (−0.67 mm yr⁻¹), CJ (−0.87 mm yr⁻¹) and SC (−0.92 mm yr⁻¹). Areas with increasing ET are mainly located along the eastern coastal belt and the margins of the southwestern mountainous regions, whereas the central inland parts of the subtropical zone generally exhibit weak, non-significant declines.

In addition, the multi-year mean annual ET exhibits a distinct pattern of higher values in the south than in the north, and higher values in coastal than in inland regions (Figure 3b). Subregions located along the southeastern coast, such as TB (954.13 mm), DN (849.94 mm) and MY (761.38 mm), have the highest ET, indicating strong evaporative demand. By contrast, subregions situated mainly in the northwestern inland areas, including SC (549.92 mm), GZ (571.45 mm) and HJ (579.27 mm), show relatively low annual ET.

From 2001 to 2020, the per-pixel probability density function (PDF) of annual total ET in subtropical China spans a range of 383–1152 mm, with a relatively sharp peak occurring in each year (Figure 3d). This indicates that although annual ET exhibits large spatial heterogeneity across the study region, its values tend to fall within a relatively stable range, with few occurrences of extreme fluctuations (Figure 3d). Over 2001–2020, the ET value corresponding to the maximum probability density consistently lies between 615 and 650 mm, which likely reflects a typical ET level for the subtropical zone. In addition, although interannual differences in the shape of the PDFs are small, the overall range of the per-pixel annual ET distribution has gradually broadened over time. In particular, after 2010 both the extent and probability of higher-ET areas increased, which is consistent with the finding that regional annual ET has been increasing, with an accelerated rate of change after 2010.

To further investigate the spatial characteristics of ET, we analysed the pixel-scale spatial patterns of annual total ET and its trends (Figure 4), providing a basis for subsequent examination of the driving factors and regional differences. During 2001–2020, high annual ET values are mainly concentrated in the western TB subregion, central–western DN, and southern XM, whereas low ET values are primarily found in central GZ, western SC and northern XM (Figure 4a). The spatial distributions of pixel-scale ET trends (slope values) are shown in Figure 4b,c. Overall, 64.53% of pixels exhibit positive slopes in annual ET. Among these, 29.23% show a significant increasing trend (p < 0.01) and 11.04% show a marked increasing trend (p < 0.05) (Figure 4c). TB, DN, XM and YN are characterised by significant increasing ET with significant trends. In contrast, pixels with significant decreasing trends (p < 0.01), accounting for 7.03% of the study area, are mainly distributed around major urban agglomerations. This pattern is likely related to land-use changes associated with urban expansion, particularly the conversion of agricultural land to built-up areas during rapid urbanisation.

Based on the Hurst exponent calculated from annual total ET during 2001–2020, the spatial distribution of future ET trends in subtropical China is shown in Figure 4d. Pixels characterised by persistent decrease, anti-persistent decrease (i.e., past decreases in ET are likely to be followed by increases), anti-persistent increase (i.e., past increases are likely to be followed by decreases) and persistent increase account for 13.14%, 21.14%, 40.19% and 25.53% of the study area, respectively. Overall, areas where ET is likely to decline in the future (53.33%) are broadly comparable in extent to those where ET is likely to increase (46.67%). Nevertheless, the mean slope of annual ET in areas projected to experience future decreases (−2.51 mm yr⁻¹) is much smaller in magnitude than that in areas projected to experience future increases (4.43 mm yr⁻¹), suggesting a tendency toward an overall increase in total ET in subtropical China under the assumption of stationary climatic and anthropogenic conditions. It should be noted, however, that Hurst-based projections do not account for future changes in climate drivers or human activities, and the observed tendencies should be interpreted with caution. TB, DN, XM and YN are dominated by persistent increases in ET, and this pattern closely matches the spatial distribution of high multi-year mean ET. This indicates that in these high-ET areas, evaporative capacity has continued to strengthen over time, leading to a positive relationship between ET magnitude and ET persistence.

3.2. Impacts of Climate Change and Human Activities on ET

3.2.1. Characterising the Feature Importance of Driving Factors

The correlation analysis between ET and these variables (Figure 5a) shows that, across subtropical China, most factors are positively correlated with ET. Among them, SWDown (r = 0.56) and LAI (r = 0.50) exert the strongest influences, where SWDown controls the surface energy input, while LAI reflects the transpiration capacity of vegetation, and both play particularly important roles in regulating ET under the generally favourable hydrothermal conditions of the subtropical zone. It is also noteworthy that Tair, VAP and Qair are highly correlated with each other (r > 0.80), indicating that under the subtropical monsoon climate these meteorological variables may jointly represent the background climatic field governing regional evaporative demand.

Figure 5b illustrates the relationships between ET and its driving factors captured by the XGBoost–SHAP framework. The SHAP scatter plots show both the magnitude and direction of each predictor’s influence on the model output. On the x-axis, SHAP values indicate the contribution of each climatic or vegetation factor to changes in ET, while each point represents a sample, coloured from red to blue to denote high to low predictor values, respectively. The factors are ordered from top to bottom according to their overall importance for ET. SWDown and LAI emerge as the two most influential variables: as their SHAP values transition from negative to positive, the point colours shift from blue to red, indicating a strong positive effect of higher SWDown and LAI on ET. LAI exhibits a particularly wide range of SHAP values, with especially large spread under high-LAI conditions, implying a strongly nonlinear and highly variable influence on ET. By contrast, SWDown shows larger SHAP variability at low values, suggesting that ET is very sensitive to changes in radiation under energy-limited conditions. To further investigate the joint effects of SWDown and LAI on ET trends, mean ET was stratified for different combinations of SWDown and LAI for 2001–2010, 2011–2020 and 2001–2020 (Figure 5c). Overall, ET increases markedly with rising SWDown and LAI, and the joint distribution of these two variables progressively shifts towards higher values over time. In areas with LAI > 4 and SWDown > 180 W m⁻², ET during 2011–2020 is clearly higher than during 2001–2010.

The SHAP scatter plots also indicate that Tair exerts a positive influence on ET in subtropical China, whereas Soil_w tends to have a negative effect, consistent with the correlation analysis. Although VAP does not show a pronounced positive or negative tendency, it remains one of the more important climatic controls on ET. By contrast, Soil_w, Precip and Qair have relatively low importance scores in the model. In particular, Qair does affect ET, but its strong correlations with Tair and VAP (r > 0.80) mean that its independent contribution is limited, leading to low apparent importance. Based on the SHAP importance threshold (SHAP > 10), five key predictors were ultimately selected to construct the multiple linear regression model used to predict ET: SWDown, LAI, Tair, VAP and Wind_sp. To ensure that multicollinearity among these predictors would not adversely affect the regression results, variance inflation factors (VIFs) were calculated for all five variables. The results indicated that all VIF values were below 10, suggesting that multicollinearity is not likely to substantially bias the model coefficients or the interpretation of predictor contributions.

3.2.2. Spatiotemporal Variations in the Key Drivers

In the regional statistical analysis, the ten subregions were grouped according to their mean latitude into a southern group (red) and a northern group (blue). As shown in Figure 6a, among the five key drivers, SWDown exhibits the largest interannual variability, whereas Tair and VAP show relatively small fluctuations and a consistent ordering between the two groups, further confirming their strong mutual correlation (Figure 6a). The highest mean values of LAI, Tair, VAP and Wind_sp occur in TB, MY, MY and TB, respectively, while the lowest means are found in SC, XM, XM and SC.

We then performed group-wise regressions between ET and the main climatic and vegetation factors (Figure 6b). Except for SWDown, the distributions of the key drivers are generally more compact in the northern group, whereas their ranges are broader in the southern group. The regression results indicate that ET in the northern group responds strongly to SWDown, with a steeper slope than in the south. In contrast, ET in the southern group is most strongly correlated with LAI and exhibits a higher LAI–ET slope. This implies that SWDown has greater sensitivity and explanatory power for ET in the northern group, while LAI plays a more dominant role in the south. Overall, the relationships between ET and Tair, VAP and Wind_sp are weaker and more spatially variable, providing secondary explanatory power; this is consistent with the SHAP scatter plots in Figure 6b.

To further explore the spatial heterogeneity of driver effects on ET, we examined the spatial correlations between ET and the main predictors across subtropical China (Figure 6c). Overall, both SWDown and LAI are positively correlated with ET. SWDown–ET correlations are generally higher and more spatially continuous in the northern group, whereas in southern subregions such as DN and XM, they display pronounced spatial heterogeneity, likely due to stronger influences of topography, cloud cover and local climatic disturbances. LAI is significantly and positively correlated with ET over most areas, with correlation coefficients exceeding 0.5 across large regions and reaching up to 0.7 in some zones. VAP and Wind_sp show clear regional differentiation: positive VAP–ET correlations are concentrated mainly in the northern group, while positive Wind_sp–ET correlations occur primarily in the southern group. Notably, in TB, ET is negatively correlated with SWDown and VAP but positively correlated with LAI, Tair and Wind_sp, which is consistent with a humid island environment characterised by ample moisture and limited radiation.

3.2.3. Spatial Distribution of Dominant Drivers of ET Changes

Building on the multiple regression model and the spatial pattern of climate-dominated factors, we overlaid the significance of the ET residual trends and extracted, for each pixel, the factor with the largest relative contribution to identify the dominant drivers of ET changes during 2001–2020 (Figure 7). Overall, spatial differences in ET across subtropical China are mainly controlled jointly by meteorological and vegetation factors. SWDown and LAI emerge as the most important and explanatory variables over most of the region, accounting for 31.36% and 31.07% of the total area, respectively, and thus represent the key controls on regional ET changes (Figure 7a). In particular, LAI dominates 50.29% of the area in DN, whereas SWDown dominates 65% of the area in HJ.

Figure 7b presents a radar chart of the relative area contributions of each dominant factor. ET processes exhibit pronounced spatial differentiation among the subregions. In the northern group (blue semicircle), ET variations are primarily dominated by SWDown, with LAI acting as an important secondary regulator in most areas. In the southern group (red semicircle), ET changes are jointly dominated by LAI and Tair. As shown in

Table 1. Relative contributions of dominant drivers to ET changes across ten subtropical subregions.

Dominant Drivers	Northern Subregions					Southern Subregions
	CJ	HJ	SC	GZ	JN	YN	XM	MY	TB	DN
LAI	27.66%	18.32%	19.27%	24.47%	43.37%	36.61%	12.16%	25.03%	12.96%	50.29%
SWDown	29.67%	65.00%	55.63%	46.27%	21.80%	22.11%	5.24%	25.01%	10.01%	8.89%
VAP	23.39%	3.28%	15.46%	20.62%	17.24%	11.40%	14.17%	19.27%	22.45%	6.97%
Wind_sp	6.47%	3.13%	2.52%	4.93%	4.16%	2.32%	8.47%	4.22%	1.51%	4.29%
Tair	8.15%	8.27%	2.72%	1.99%	5.31%	11.90%	30.72%	19.68%	33.75%	16.50%
Residual ET	4.66%	2.00%	4.40%	1.72%	8.12%	15.66%	29.24%	6.79%	19.33%	13.06%

, the mean contribution of SWDown is high in the north (43.67%) and shows large spatial variability, whereas in the south it is lower (14.25%) and less variable. By contrast, the contribution of LAI is relatively stable across both groups. Among the dominant drivers, SWDown contributes the most in HJ, reaching 65%, while LAI dominates in DN, accounting for 50.29%. Although the dominant areas of the other factors are smaller, Tair controls ET changes over 33.75% of TB and 30.72% of XM, respectively. VAP is dominant over 23.39% of CJ and 22.45% of TB, while Wind_sp dominates 8.47% of XM. Residual ET exhibits relatively large contributions in XM, TB, and YN, accounting for 29.24%, 19.33%, and 15.66% of their areas, respectively.

Some areas are strongly disturbed by human activities, making it difficult for the model to accurately predict ET changes. Comparison with the 2020 land-use map (Figure 1) shows that pixels classified as “Residual ET” overlap substantially with built-up land, cropland and related categories, indicating that ET in these areas is heavily influenced by human activities, particularly land-use changes associated with urbanisation. Nevertheless, unmodeled natural processes, such as deep soil moisture, groundwater, fog inputs, and CO₂ fertilisation, may also contribute to residual ET trends, potentially leading to an overestimation of anthropogenic effects. These effect are particularly pronounced in karst and mountainous regions, where residual trends are more likely dominated by unmodeled natural processes. In XM and TB, where terrain is highly complex with strong topographic relief, the contribution of residual ET may be influenced by local microclimatic variability, slope and aspect effects, and spatial heterogeneity in soil moisture. In YN, where extensive karst landforms exist, the large residual ET contribution reflects the influence of hydrological heterogeneity, including groundwater and subsurface water flows, rather than solely human intervention.

4. Discussion

4.1. Accuracy Assessment of ET Datasets

Evaluating remote-sensing ET products against ground-based flux observations is not only a key means of testing their applicability across different regions, but also a prerequisite for deepening our understanding of ET processes and improving the accuracy of regional water-cycle simulations [8]. Previous studies have shown that different ET products differ substantially in their retrieval algorithms, input variables and parameterisation schemes, which can introduce systematic biases into the ET estimates [5,45]. In this study, we selected two ET products with consistent spatial and temporal resolutions, MOD16 and PML-V2, and validated their grid-based ET against flux-tower observations to assess their accuracy and suitability for subtropical China. The results indicate that MOD16 yields significantly higher annual ET totals than PML-V2 in this region, whereas PML-V2 exhibits clearly better regional applicability and accuracy (Figure 2).

Both MOD16 and PML-V2 are based on Penman–Monteith-type formulations, but they differ substantially in model structure, input data, and process representation. As shown in

Table 2. Comparison of input datasets for the MOD16 and PML-V2 evapotranspiration products.

Data Type	MOD16 (ET)	PML-V2 (Ec, Es, Ei)
Meteorological Data	GMAO-MERRA (1° × 1.25°) $T_a,P_{air},RH,R_{s\downarrow }$	GLDAS_2.1 (0.25° × 0.25°) $Prcp,T_a,P_{air},R_{s\downarrow },R_{L\downarrow },WS$
Land Cover Product	MOD12Q1-UMD (1 km/a)	MCD12Q1-IGBP (500 m/a)
LAI	MOD15A2 (1 km/8 d)	MCD15A3H (500 m/8 d)
Albedo Data	MOD43C1 (0.05°/16 d)	MCD43A3 (500 m/8 d)
Surface Emissivity	-	MOD11A2 (500 m/8 d)
CO2 Concentration Data	-	NOAA-GAMS/MMD (~1°)

, compared with MOD16, which uses GMAO-MERRA meteorological forcing at 1° × 1.25° and 1 km LAI with coarser albedo and 1 km land cover, PML-V2 employs higher-resolution meteorological forcing including additional variables such as downwelling longwave radiation and wind speed, 500 m LAI and 8-day albedo and emissivity, and 500 m land cover. PML-V2 uses the Leuning surface conductance model together with the Penman–Monteith equation, and incorporates global CO₂ concentrations to simulate stomatal behaviour, linking transpiration with photosynthetic activity at the canopy scale. This explicit water–carbon coupling allows ET to respond physiologically to variations in GPP and atmospheric CO₂, improving the representation of ecosystem water–carbon interactions compared to MOD16, which treats ET and GPP independently [7,46,47]. Numerous earlier studies likewise report that PML-V2 outperforms MOD16 in humid parts of mainland China, and that MOD16 tends to overestimate ET in southern Chinese basins [48,49]. Pan et al. (2022) further showed that, among several ET products used to improve hydrological model simulations in humid East China, PML-V2 provided the greatest performance gains, which indirectly corroborates the superior performance of PML-V2 found in this study [50]. In addition, Chen et al. (2022) showed that, in the Lancang–Mekong River basin, PML-V2 performs best for evergreen broadleaf forest (EBF) areas, whereas MOD16 exhibits the largest errors among the evaluated products [51]. These findings help explain why, in subtropical China, PML-V2 exhibits larger seasonal peak amplitudes and higher regional accuracy and applicability than MOD16.

The three flux sites exhibit distinct intra-annual and interannual patterns of ET (Figure 2b). In terms of intra-annual variability, the transpiration fraction (T/ET) at QYZ is high and peaks in summer, and ET there responds strongly to net radiation and vapour pressure deficit. In addition, the decline in soil moisture during midsummer strengthens the control of atmospheric dryness on stomatal conductance, thereby amplifying the seasonal peak–trough contrast in ET [52,53]. For interannual variability, QYZ is a transpiration-dominated system that is highly sensitive to phenology and to the onset and retreat of the monsoon, so ecosystem-side interannual differences tend to be amplified [54]. By contrast, DHS and XSBN represent evergreen, energy-limited systems with strong canopy interception and fog–water inputs, where ecohydrological buffering tends to damp both intra-annual and interannual variability in ET [55,56]. This likely explains the larger interannual variability at QYZ and the smaller variability at DHS and XSBN found in this study.

4.2. Spatiotemporal Variability and Persistence of ET

In this study, the mean annual ET over subtropical China estimated using the PML-V2 product is 669 mm, with a clear decreasing gradient from the humid southeastern coastal region toward the drier northwestern interior. Similar spatial patterns and long-term trends have also been reported by Fu et al. (2022), Cheng et al. (2021) and Li et al. (2018) [9,57,58]. Although few studies have applied PML-V2 to estimate ET across the entire subtropical region of China, the magnitude obtained in this study falls within the range reported for comparable subtropical subregions using the same product, such as southwestern China (703 mm), the Yangtze River Basin (593 mm) and the Three Gorges Reservoir area (585 mm) [59,60,61]. As shown in Figure 3a, annual total ET over subtropical China exhibits a significant increasing trend during the study period (2.03 mm yr⁻¹, p < 0.01), with a notably accelerated increase after 2010. Since the early 2000s, widespread vegetation greening across China has contributed to an overall upward trend in ET [62]. Subsequently, since 2013, effective air-pollution control measures have substantially reduced the negative radiative effects of aerosols, leading to surface “brightening” and increased available energy, which further accelerated the rise in ET [63,64]. It is worth noting that the pronounced interannual minimum observed in 2010 is likely associated with the severe drought triggered by the 2009–2010 Central Pacific El Niño event [65,66]. This drought may have temporarily suppressed ET by reducing precipitation and soil moisture availability and by increasing atmospheric water stress, thereby limiting vegetation transpiration. After the drought, the gradual replenishment of soil moisture and recovery of vegetation activity could have induced a lagged rebound in ET. Therefore, the apparent acceleration of ET after 2010 may partly reflect a post-drought recovery signal superimposed on the long-term effects of vegetation greening and increased surface radiation, rather than an abrupt change driven by a single factor.

Despite the overall increasing trend in ET, pronounced spatial heterogeneity exists across subtropical China, with localised regions exhibiting declining ET trends. Over the past few decades, increases in terrestrial ET have been driven primarily by the combined effects of global warming [16,67] and vegetation greening [68]. In the karst regions of southwestern China, large-scale ecological restoration projects such as natural forest protection [69,70], the Grain for Green Programme [71] and rocky desertification control [72] have led to significant increases in ET. On the southern flank of the Himalayas, warming-induced vegetation greening [73] and earlier snowmelt, which lengthens the growing season [74], have also enhanced ET. The increase in ET over the Taipei region is closely linked not only to afforestation policies [75], but also to air-pollution control measures that have reduced the negative radiative forcing of aerosols, thereby further raising actual ET [76,77]. In contrast, a significant decline in ET has been observed in the urban–rural transition zones of the Yangtze River Delta. Previous studies have reported that, in the Yangtze River Delta, the conversion of paddy fields to urban land reduces ET more strongly than climatic factors, with the urban–rural fringe showing the most marked suppression of ET [73]. In the Sichuan Basin and the upper–middle Han River region, ET dynamics are influenced by the expansion of the Chengdu–Chongqing metropolitan area [78], the increasing frequency of regional drought events [79], protection of the Danjiangkou drinking-water source area [80] and regulation policies associated with the South-to-North Water Diversion Project [80]. These findings are consistent with the spatial patterns of annual ET trends identified in this study (Figure 4b,c).

4.3. Attribution of ET Changes

ET is primarily controlled by radiation conditions and vegetation–soil characteristics, with solar radiation and LAI among the key regulating factors [23,81,82]. Our results are consistent with this understanding (Figure 5 and Figure 7). Similar increasing ET trends and related driving mechanisms have also been reported in other humid and subtropical regions. For example, Kramer et al. (2015) found that ET increased over the eastern United States during the twentieth century, largely associated with increasing precipitation and possible vegetation greening [83]. A comparable pattern has also been reported at the national scale in China, where long-term ET trends are mainly driven by vegetation and precipitation changes [9,84]. In contrast, precipitation shows relatively low explanatory power for interannual ET variability in subtropical China. Our results instead reveal a distinct north–south divergence in ET controls, with northern subregions mainly constrained by downward shortwave radiation and southern subregions jointly regulated by LAI and air temperature. Evidence from Japanese forest catchments also indicates that annual forest ET is closely linked to air temperature, supporting our finding that temperature and vegetation jointly regulate ET in humid southern subregions [85].

The southern subregions generally exhibit higher SWDown and LAI than the northern subregions, suggesting more abundant energy supply and stronger vegetation activity (

Table 3. Comparison of hydroclimatic and vegetation characteristics between northern and southern subregions during 2001–2020.

Variable	Northern Subregions					Southern Subregions
	CJ	HJ	SC	GZ	JN	YN	XM	MY	TB	DN
LAI	1.17	1.57	1.02	1.59	2.04	1.85	1.62	1.83	3.23	2.39
SWDown	173.67	156.57	138.30	154.92	171.31	166.72	181.39	178.51	179.67	186.06
VAP	1.52	1.22	1.57	1.42	1.65	1.12	0.94	2.00	1.81	1.58
Wind_sp	2.47	1.90	1.32	1.81	2.10	2.04	2.10	2.05	3.60	1.60
Tair	16.67	13.85	17.28	15.53	17.94	13.21	7.75	21.43	18.70	18.45

). However, the contribution of SWDown to ET changes is much lower in the south, indicating that ET is less constrained by radiation availability under energy-sufficient conditions and becomes more sensitive to vegetation-related processes [86]. By contrast, the lower SWDown and LAI in the northern subregions imply weaker energy input and more limited vegetation activity, making ET more directly constrained by variations in SWDown [87]. Differences in vegetation functional types and canopy structures may further reinforce this contrast. Northern subregions are mainly characterised by cropland, mixed forests, and plantations with relatively lower LAI and stronger seasonal vegetation activity, whereas southern subregions include more evergreen broadleaf forests, tropical seasonal forests, mountainous forests, and managed plantations with higher LAI, more complex canopy structures, and longer growing seasons. In addition, complex terrain in southern regions such as XM and YN enhances spatial heterogeneity in thermal and moisture conditions. Together, these differences suggest that ET in the north is more radiation-limited, whereas ET in the south is jointly regulated by LAI and air temperature.

4.4. Limitations and Future Perspective

In contrast to earlier nationwide attribution studies that offer only a single ranking of drivers, our pixel-wise identification of dominant factors and north–south subregional comparison reveal a gradual shift from a radiation-limited to a vegetation-limited ET regime within subtropical China, thereby enhancing understanding of ET drivers at a finer spatial scale. Nevertheless, this north–south divergence should be further tested using hydroclimatic indicators such as aridity index, radiation regime, potential evapotranspiration, or water–energy limitation metrics [88].

Although significant progress has been made, the quantitative assessment of human impacts on ET remains an unresolved scientific challenge [89]. In this study, the residual component was compared with static land-use patterns, but temporal dynamics such as urban expansion, cropland conversion, and ecological restoration were not explicitly incorporated. Future work should therefore combine multi-temporal land-use data and urbanisation indicators to better quantify the dynamic relationship between land-use change and residual ET trends [30]. In addition, the complex pattern of dominant drivers in the XM subregion indicates that current models still have limitations in representing coupled human–environment processes in mountainous areas or along plateau margins with highly dissected terrain. Moreover, meteorological variables were derived from multiple datasets with different spatial resolutions and modelling assumptions. Although all variables were resampled to a common 1 km grid, resampling cannot fully remove inconsistencies among datasets. Future work should explicitly evaluate and reduce cross-dataset inconsistencies in meteorological forcing when conducting ET attribution [90].

5. Conclusions

In this study, we first compared the performance of two remote-sensing ET products, MOD16 and PML-V2, and found that PML-V2 has higher accuracy. We therefore used PML-V2 to analyse the spatiotemporal variability and drivers of ET in subtropical China. The main findings are as follows:

(1)

Over the past two decades, annual total ET in subtropical China has shown a significant increasing trend (p < 0.01), with a marked acceleration occurring after 2010. The multi-year mean annual ET exhibits a clear south–north and coast–inland gradient, with higher values in the south and along the southeastern coast and lower values in the north and inland regions. Although the area where ET is likely to decrease in the future (53.33%) is slightly larger than the area where ET is likely to increase (46.67%), the mean rate of change in the decreasing regions (−2.51 mm yr⁻¹) is much smaller in magnitude than that in the increasing regions (4.43 mm yr⁻¹). Overall, total ET in subtropical China tends to increase under the assumption of stationary climate and land-use conditions.

(2)

SWDown and LAI are identified as the key controls on ET variability in subtropical China, with strong dominance and explanatory power in the models. The strong correlations among Tair, VAP and Qair (r > 0.80) indicate that these variables jointly represent the background climatic conditions that underpin regional evaporative demand.

(3)

ET changes in the northern group of subregions are mainly dominated by SWDown, whereas in the southern group they are jointly controlled by LAI and Tair. The mean contribution of SWDown is higher in the northern group (43.67%) and exhibits larger spatial variability, but is lower (14.25%) and less variable in the southern group. In contrast, the contribution of LAI is relatively stable across both groups.

(4)

Pixels classified as “Residual ET” show strong spatial overlap with built-up land, cropland and related land-use types, indicating that ET in these areas is strongly affected by human activities, particularly land-use changes associated with urbanisation. While the residual ET component provides insight into additional non-climatic influences, its interpretation must account for the limitations of unmodeled natural processes in different subregions.