The Loess Plateau (roughly within 34°~41°N, 98°~114°E), with an area of ~632,520 km², which accounts for 6.3% of China’s total land area (Jiang et al., 2022). The average annual precipitation ranges from 200 mm in the northwest to 800 mm in the southeast, with 60%~70% of the annual precipitation occurring between June and September in the form of high-intensity rainstorms, which often cause extreme soil erosion and drought frequently occurs (Liang et al., 2020). The average annual temperature is 4°C in the northwest and 14°C in the southeast (Peng et al., 2018). Consequently, the region is characterized as a typical transitional zone from an arid to a semi-humid climate, significantly influenced by the East Asian monsoon. The predominant land cover types include grassland, cropland, and forest ( Figure 1 ).

Based on the ecological restoration areas and geographical characteristics of the Loess Plateau, and following the division methods used in previous studies (Jiang et al., 2022; Yang Y. F. et al., 2019), the Loess Plateau is divided into five sub-regions: the Northern Loess Plateau (NLP), the Middle Loess Plateau (MLP), the Eastern Loess Plateau (ELP), the Western Loess Plateau (WLP), and the Southern Loess Plateau (SLP) ( Figure 1 ). Among these, the NLP is a typical semi-arid area with a mean annual precipitation below 400 mm, mostly covered by deserts or desertified land. The MLP is the primary region for vegetation restoration projects (Jiang et al., 2022), while the WLP consists mainly of high tableland and gully regions. The SLP and ELP are primarily composed of broad valley plains and rocky mountain areas.

Climatologically, the total growing season ET on the Loess Plateau exhibited a decreasing pattern from southeast to northwest (from > 2 mm/day in the southeast to < 1 mm/day in the northwest), closely mirroring the precipitation pattern and reflecting the limiting role of water resources on the ET process in this semi-humid to semi-arid region ( Figure 2A ). Among the three components of ET, Ec was the primary contributing component, accounting for ~70% of total ET from 2000 to 2022. Similar to ET, the spatial distribution of Ec ( Figure 2B ) also showed a decreasing trend from southeast to northwest, with high-value areas exceeding 1.4 mm/day primarily located in semi-humid, forested regions, while low-value areas (<0.6 mm/day) were found in arid to semi-arid regions with precipitation below 200 mm (excluding the Hetao Plain). On the other hand, Ei was substantially lower than Ec ( Figure 2C ), with rates generally below 0.1 mm/day outside forested areas, contributing minimally to total ET, with values ranging from 3% to 6%.

In contrast to the two vegetation-related ET components (Ec and Ei), the distribution pattern of Es was characterized by low values in the southeast and high values in the northwest ( Figure 2D ). As the second largest contributor to total ET (~19% to 32%), relatively higher Es values (~1 mm/day) were observed in regions with annual precipitation below 400 mm, where lower vegetation cover allowed Es to dominate the ET process. In comparison, Es in densely vegetated areas was weaker, generally below 0.4 mm/day, indicating a competitive relationship between vegetation water consumption and soil evaporation.

From 2000 to 2022, the regional mean ET time series of the Loess Plateau showed a significant increasing trend (p < 0.05) ( Figure 3A ), with an increase rate of 0.164 mm/day per decade (R²=0.5197). Both vegetation-related components, Ec and Ei, also showed significant increasing trends (p < 0.05) with respective increase rates of 0.194 and 0.019 mm/day per decade ( Figures 3B, C ). In contrast, the regional mean Es demonstrated a significant decreasing trend (p < 0.05) over the same period ( Figure 3D ), with a decrease rate of −0.043 mm/day per decade, partially offsetting the increasing trends of the two vegetation-related components. As a result, the total ET showed a relatively smaller increasing trend compared to Ec.

The distribution of ET change trends indicated that the majority of the Loess Plateau experienced a significant increase in ET from 2000 to 2022, with the exception of marginal areas in the NLP, WLP, SLP, and ELP ( Figure 4A ). Compared to ET, Ec exhibited a broader area with a significant increasing trend ( Figure 4B ), particularly in the northwestern and southeastern corners of the Loess Plateau. In the MLP, where substantial vegetation recovery occurred, the trend in Ec exceeded that of ET (> 0.5 mm/day per decade). The water intercepted by the vegetation canopy showed a marked increase in the southeastern semi-humid region of the Loess Plateau (MLP, WLP, and ELP) ( Figure 4C ), with particularly pronounced increasing trends (> 0.1 mm/day per decade) in forested areas. In contrast, Es displayed opposite change trends between the northwestern and southeastern parts of the Loess Plateau ( Figure 4D ). Most of the Loess Plateau exhibited significant decreasing trends in Es, particularly in the semi-humid MLP, WLP, and ELP, while a minority of areas (primarily in the semi-arid and arid NLP and northern WLP) showed significant increasing trends in Es.

The regional mean NDVI time series of the Loess Plateau exhibited a significant (p < 0.001) increasing trend from 2000 to 2022 ( Figure 5A ), with an increase rate of 0.045 per decade, indicating an overall greening trend for the Loess Plateau since the implementation of large-scale ecological projects. The spatial pattern of NDVI trends revealed that almost the entire Loess Plateau experienced significant increases in NDVI, particularly in the MLP and WLP regions, where the NDVI increase exceeded 0.1 per decade ( Figure 6A ). During the same period, the Loess Plateau as a whole did not exhibit significant trends in either temperature or precipitation during the growing season ( Figures 5B, C ), with change rates of 0.132°C per decade and −0.084 mm/day per decade, respectively. Moreover, the spatial distribution of trends in each pixel did not show significant trends in most areas of the Loess Plateau ( Figures 6B, C ). Most of the Loess Plateau experienced a rise in temperature from 2000 to 2022, with a declining trend observed only in the eastern WLP and western SLP ( Figure 6B ). Additionally, there were decreasing trends in precipitation across the majority of the Loess Plateau ( Figure 6C ), with only parts of the eastern MLP and southern WLP showing non-significant increases in precipitation.

By employing the methodology proposed by Huang, we assessed whether vegetation greening is the direct cause of the trends in ET on the Loess Plateau. Among the pixels showing a significant trend in ET (covering approximately 71.4% of the total area of the Loess Plateau), around 82.7% also experienced a significant increase in NDVI ( Figure 7A ), while only about 17.3% did not show any significant change in NDVI. This result indicates that the majority of areas with significant ET changes were directly caused by vegetation greening. Compared to ET, the area with a significant increase in Ec was larger, covering approximately 75.1% of the Loess Plateau ( Figure 7B ). In areas with significant increases in Ec, over 81.4% of grid points also showed a notable increase in NDVI, suggesting that the rise in Ec is predominantly due to vegetation recovery. Less than 20% of the grid points, primarily located in arid and semi-arid zones, exhibited no significant change in vegetation. There were few pixels with significant decreases in Ec, mainly associated with changes in vegetation cover. Furthermore, approximately 80.5% of all pixels showing a significant increase in Ei also experienced significant rises in vegetation ( Figure 7C ), indicating that the increase in Ei is mainly driven by vegetation changes. Less than 15% of these areas (primarily in the western WLP and the cropland areas of SLP and ELP) showed Ei enhancement driven by climatic changes. Of the small number of pixels with a significant decrease in Ei in arid and semi-arid regions, about one-fourth of these areas were attributed to vegetation changes.

Between 2000 and 2022, the area with a significant decreasing trend in Es covered approximately 56.7% of the total area of the Loess Plateau, primarily distributed across the MLP, ELP, eastern WLP, and parts of the SLP. In these areas, the majority (about 77%) coincided with a significant increase in NDVI ( Figure 7D ), indicating that the decrease in Es was related to changes in vegetation cover. In addition, about 12.9% of the total area showed a significant increase in Es, sporadically located in arid and semi-arid regions, with approximately 65% of these areas also driven by changes in vegetation cover. Although Es was not directly produced by vegetation, long-term vegetation changes remained a primary cause of the changing trends in Es.

To explore the impact of vegetation conditions and climate factors on ET at an interannual time scale, Pearson correlations between ET (and its components) and NDVI (or climatic factors) were calculated for each pixel after removing their corresponding trend components ( Figure 8 ). The results revealed that ET and Ec were significantly (p < 0.001) and positively correlated with NDVI in some areas of the MLP, NLP, and WLP ( Figures 8A, B ). In contrast, the positive correlation of Ei and the negative correlation of Es were largely insignificant ( Figures 8C, D ). In comparison, the correlations of ET (Ec and Es) with the two climatic factors were much stronger, particularly in the arid and semi-arid regions (p < 0.001, Figures 8E–G, I–K ). For instance, there were significantly negative correlations between temperature and ET/Ec/Es, suggesting that in these water-limited areas, increased temperatures and associated drought stress may inhibit the ET process. Additionally, a significantly positive relationship was found between precipitation and ET/Ec/Es over large areas in the northwest, although some negative correlations were observed in the southeastern marginal regions. The correlations between precipitation and temperature indicated that interannual variations in ET were predominantly controlled by water availability in these water-limited regions, rather than energy availability. These findings also suggest that climate is the primary factor influencing the interannual variations in ET and its components in the arid and semi-arid regions of the Loess Plateau.

To verify the robustness of our results, the China meteorological forcing dataset (CMFD) (Yang K. et al., 2019) were also conducted correlation tests with ET with long term trends excluded, this dataset integrates ground observations and satellite data and are widely proven to have good performance in representing the precipitation and temperature pattern (Yang et al., 2010; He et al., 2020). The results of CMFD demonstrate that the spatial patterns in the correlation of ET with both temperature and precipitation are highly consistent with that of ERA5 ( Supplementary Figures S1 ), implying that the dominant role of these climatic factors in regulating interannual variability in ET is relatively robust among different data sources. Although some researches show certain overestimation of ERA5 precipitation over the Loess Plateau (Jiang et al., 2021), our result suggests this potential overestimation of ERA5 precipitation does not undermine the validity of our conclusions.

In this study, we employed the remote-sensing-based GLEAM ET product to estimate ET changes in the Loess Plateau, as it explicitly considers the impact of soil moisture, which is a critical limiting factor for evapotranspiration in such a relatively water-limited region (Ferguson and Wood, 2010; Vinukollu et al., 2011). Generally, the GLEAM ET dataset has been proven to perform well in representing both the temporal and spatial variations of total ET and its components across China, particularly during the warm months, when compared with in situ observations and other products (Li, 2023; Yao, 2023). From the climatological results of ET over the Loess Plateau, GLEAM shows that canopy transpiration tends to be larger than soil evaporation, which contrasts with some other products (Jiang et al., 2022; Zhao et al., 2022). This difference is likely due to the following reasons. The GLEAM ET model uses a multi-layer soil model to capture soil moisture distribution and its effects on vegetation transpiration, limiting soil evaporation in dry regions while increasing Ec (Miralles et al., 2011; Martens et al., 2017). In contrast, other evapotranspiration models, such as PML and VIC models, which are based on the Penman-Monteith equation, focus primarily on vegetation and meteorological factors, emphasizing energy balance over soil moisture (Liang et al., 1994; Zhang et al., 2019). These models estimate transpiration and evaporation separately via remote sensing, potentially overestimating soil evaporation in areas with sparse vegetation. Differences in meteorological inputs also contribute to the variations in their estimates.

To determine whether long-term changes in vegetation are the direct drivers of observed trends in evapotranspiration (ET) and its components, we applied the method proposed by Huang. This approach posits that if both ET (and its components) and NDVI exhibit significant trends at the same grid point, it can be inferred that vegetation changes are the direct drivers of the trends in ET or its components, regardless of the presence of significant changes in climatic factors. Importantly, the trend in NDVI itself may be influenced by significant trends in climatic factors. However, we do not delve further into whether vegetation changes are attributable to anthropogenic activities or climatic factors, nor do we examine the extent to which climate change indirectly influences ET by altering vegetation conditions. Our analysis focuses exclusively on the direct impact of vegetation changes on ET. Conversely, when a significant trend in ET (or its components) occurs without a corresponding significant trend in NDVI, changes in climatic factors are considered the direct cause. Specifically, this method attributes vegetation changes primarily to anthropogenic activities (e.g., reforestation or land-use changes), while linking changes in climatic factors (e.g., temperature or precipitation) directly to climate change. In reality, the interactions between vegetation, climatic factors (e.g., precipitation and temperature), and ET are highly complex, involving multiple feedback mechanisms, which makes disentangling the relationships among these variables challenging. For example, in the southeastern Loess Plateau, although NDVI exhibited a significant increasing trend, the corresponding increasing trend in Ec may have been offset by the suppressive effects of significantly rising temperatures. Under such stressful conditions, plants may close their stomata, potentially explaining the overall insignificant increase in ET in these areas (Yang Y. et al., 2019). In general, Huang’s method provides a robust framework for disentangling the intricate relationships among vegetation dynamics, climatic factors, and ET trends through clear assumptions and logical classification. By avoiding the complexities of deeper causal mechanisms, this method enhances the simplicity and practicality of the analysis. Despite its simplicity, Huang’s method effectively evaluates whether vegetation changes are directly responsible for long-term trends in ET.

The growing season of the semi-arid and semi-humid Loess Plateau is characterized by relatively higher temperatures, larger vapor pressure deficits, and abundant solar radiation. Consequently, regional ET is predominantly governed by water availability in the absence of significant human interventions (Magliano et al., 2017). Theoretically, variations in ET are controlled by variations in water supply (i.e., precipitation), given the absence of large river inputs for most areas of the interior Loess Plateau (Jin et al., 2011; Zhang et al., 2017). However, extensive anthropogenic ecological restoration efforts, particularly the annual planting of new vegetation, have increased Ec, resulting in significant upward trends in ET over the past decade. This sustained increase in ET cannot be explained by non-significant trends in climatic factors but is instead attributed to vegetation greening across the plateau. The greening trend has contributed to increasing both total ET and Ec, while simultaneously being associated with a decline in soil evaporation. This decline is not only caused by increased vegetation, which leads to a soil-drying trend through additional water consumption by Ec and Ei (Allen et al., 2017), but also by the diminishing energy available as increased canopy coverage blocks a larger proportion of solar radiation from reaching the soil surface (Duveiller et al., 2018; Jiang et al., 2022).

However, when the linear trends of all variables are removed, climatic factors (precipitation and temperature) exhibit a stronger correlation with the interannual variability in ET and its components across most of the study area, compared to vegetation dynamics. As expected, the positive correlation between precipitation and ET underscores its crucial role in regulating interannual variations in ET in these semi-arid and semi-humid regions. Although rising temperatures can increase potential ET, actual ET tends to decrease. This is likely because, on the one hand, at yearly or shorter time scales, higher temperatures are often associated with lower precipitation and reduced water supply, which in turn limits ET. On the other hand, during the growing season, leaf stomatal conductance tends to close under extreme heat, leading to reduced ET. The interaction and coupling effects of vegetation changes and climate change on ET are complex. Although many studies have attempted to fully disentangle the impacts of vegetation change and climate change on ET, as vegetation also responds to terrestrial ecosystem and climate changes, few studies have focused on the different forcing mechanisms of the long-term trend and interannual variability in ET. Many studies have attempted to disentangle the impacts of vegetation and climate change on evapotranspiration (ET), but few have explored the distinct mechanisms driving its long-term trends and interannual variability. The interactions between vegetation and climate change on ET are inherently complex, as vegetation dynamically responds to changes in ecosystems and climate conditions. These intricate feedbacks among vegetation change, climate change, and ET highlight the need for further research, particularly into the differing drivers of long-term trends and interannual variability.

Although this study separately examined the relative importance of climatic factors and vegetation dynamics on long-term trends and interannual variations in ET, there are still some limitations that should be addressed in future research. First, ET exhibits high spatial heterogeneity due to its strong dependence on underlying surface conditions. However, the available ET product, with its relatively coarse resolution, was not validated against in situ ET measurements, which may introduce substantial uncertainties in the ET estimates. Second, the duration of the available data is relatively short (~20 years). While the long-term trend is significant, the attribution results may vary when using data sources with a longer temporal span. Lastly, this study only analyzed the relationship of ET with two primary climatic factors. Other factors, such as atmospheric CO2 concentration, vapor pressure deficit, and sunshine duration, may also influence ET in the Loess Plateau. Although their impacts may be relatively minor, further research is necessary to comprehensively incorporate a broader range of influencing factors in the future.