Spatiotemporal dynamics and drivers of ecological quality in Shaanxi (Yellow River Basin, China) based on optimized RSEI and SHAP

The Yellow River Basin serves as a crucial ecological barrier in China, necessitating an understanding of its ecological quality dynamics and drivers. Utilizing Moderate Resolution Imaging Spectroradiometer (MODIS) data (2000–2023), we developed an enhanced Remote Sensing Ecological Index (RSEI) for the Shaanxi section of the Yellow River Basin by applying the entropy weight method (EWM). Spatiotemporal variations were assessed using Sen’s slope and the Mann–Kendall (MK) test (Sen–MK) and the Hurst exponent (H), while driving forces were interpreted using SHapley Additive exPlanations (SHAP). The mean RSEI was 0.51, indicating ecological quality above medium with a slight upward trend. Improvement was predominant, following a “south-high, north-low” pattern; slightly and significantly improved areas comprised 45.05% and 11.17%, while significant degradation covered 6.20%. Hurst analysis showed 56.96% of areas had significant increases, with 56.64% likely to improve further. Land use changes and natural factors jointly influenced RSEI, with cropland/grassland conversion to forest enhancing ecological quality. Soil moisture emerged as the dominant climatic factor, while interactions among wind speed, precipitation, and SOIL moisture bolstered ecosystem stability. These findings offer a scientific basis for ecological management and sustainable development in the Shaanxi section of the Yellow River Basin.

1 Introduction

The ecological environment’s quality is crucial for human survival and development (Cai et al., 2023). In the past 2 decades, China’s ecological environment has generally improved (Fang et al., 2023). However, comprehensive evidence on the spatiotemporal evolution of ecological quality and its driving mechanisms remains limited at sub-basin and regional scales, where complex terrain and heterogeneous land cover can amplify coupled natural–human influences (Zhu et al., 2024). Such investigation is essential for scientifically developing governance strategies and improving regional ecological quality.

Remote sensing technologies have been extensively applied to monitor regional ecosystem dynamics. The Remote Sensing Ecological Index (RSEI) integrates greenness, wetness, dryness, and heat into a composite indicator, typically via principal component analysis (PCA), enabling spatially explicit and quantitative assessment of ecological quality. This methodology has been widely implemented in the Yellow River Basin (Wang et al., 2023b; Qin et al., 2024; Zhou et al., 2024), the Yangtze River Basin (Yang et al., 2021), and the Loess Plateau (He et al., 2024).

From the domestic and international research landscape, RSEI has been widely used for large-scale ecological monitoring and governance evaluation; however, several issues remain for ecologically heterogeneous and transitional regions. First, conventional RSEI may be affected by Normalized Difference Vegetation Index (NDVI) saturation and data-dependent PCA weighting, which can reduce robustness and comparability across space and time (Wang et al., 2023a; Liu et al., 2024a) Second, many existing studies emphasize trend detection, whereas persistence-oriented diagnosis is relatively less addressed (He et al., 2024). Third, although driver identification increasingly relies on machine-learning models, mechanism-oriented interpretation is often limited without explainable tools to reveal nonlinear responses and potential interactions among drivers (Hosseiny et al., 2022).

To address these issues, this study advances the conventional RSEI framework beyond incremental modification in three aspects: (i) we introduce the kernel normalized difference vegetation index (kNDVI) to mitigate the saturation effect inherent in traditional NDVI while accommodating complex phenological cycles (Wang et al., 2023a; Chen et al., 2025); (ii) we adopt the entropy weight method to assign indicator weights according to their information content and variability, thereby improving the stability and interpretability of index construction compared with PCA-based weighting (Liu et al., 2024b; Zhou et al., 2025). and (iii) rather than reporting trends alone, we couple multi-temporal diagnosis (Sen–MK and Hurst) with explainable driver attribution (RFR–SHAP), enabling an integrated monitoring–persistence–attribution chain for ecologically transitional regions. Detailed procedures are provided in Supplementary Notes S1.

Beyond ecological index construction, trend identification and driving factor analysis remain essential components. The Sen–MK test is widely utilized for long-term trend analysis due to its robustness and outlier insensitivity (Liu et al., 2023; He et al., 2024; Chen et al., 2025). The Hurst exponent measures time-series persistence characteristics and is commonly used to infer long-term ecosystem evolution tendencies (He et al., 2024). Machine-learning methods have gained prominence for identifying driving factors, but they are often considered “black boxes.” To improve interpretability, SHapley Additive exPlanations (SHAP) based explainability has been increasingly adopted to quantify feature contributions and reveal nonlinear relationships, thereby strengthening confidence in driver attribution (Hosseiny et al., 2022; Antonini et al., 2024; Meng et al., 2024).

Overall, this study investigates the spatiotemporal dynamics and drivers of ecological quality in the Shaanxi section of the Yellow River Basin during 2000–2023 with three explicit objectives: (1) to construct an optimized RSEI by integrating kNDVI and entropy-based weighting; (2) to diagnose ecological change using trend detection and persistence analysis (Sen–MK and Hurst exponent); and (3) to identify dominant drivers and their nonlinear effects using a machine-learning model coupled with SHAP-based interpretation. By explicitly addressing NDVI saturation, enhancing weight stability for multi-temporal comparison, and integrating interpretable driver attribution, this framework provides methodological advancement for ecological assessment in heterogeneous sub-basin regions rather than an incremental index adjustment. These efforts provide an empirical basis for targeted ecological management and restoration planning in this ecologically transitional region.

2 Data and methods

2.1 Study area overview

The Shaanxi section of the Yellow River Basin is located in the core area of the middle reaches of the Yellow River, spanning 105°30′–111°15′E and 34°15′–39°35′N. It covers several municipalities, including Weinan, Tongchuan, Xi’an, Xianyang, Yan’an, Yulin, and Shangluo. The Yellow River enters Shaanxi from Fugu County, Yulin City, flows through the province for 719 km, and exits at Tongguan County, Weinan City. As shown in Figure 1, the study area covers approximately 133,000 km², accounting for 64.5% of the total area of Shaanxi Province. It lies in the transitional zone between the Loess Plateau and the Qinling Mountains, with significant topographic variation and complex ecological types. The topography shows a stepped pattern with higher elevations in the north and south and lower elevations in the central part. The northern part comprises the Northern Shaanxi Loess Plateau, the central part is the Guanzhong Basin, and the southern part consists of the Qinling Mountains.

Figure 1

Figure 1. Location and topographic map of the study area.

2.2 Data sources and preprocessing

2.2.1 Remote sensing data processing

This study utilized standard MODIS products on the Google Earth Engine (GEE) platform. These datasets offer broad spatial coverage, high temporal resolution, and consistent image quality, making them widely applicable for regional-scale ecological monitoring and assessment (Liao and Jiang, 2020; Zhang et al., 2022; Zhu et al., 2024). To capture typical ecological conditions, MODIS standard products from the vegetation growing season (June to September) between 2000 and 2023 in the Shaanxi section of the Yellow River Basin were selected, with a spatial resolution of 500 m. We conducted image preprocessing on the GEE platform, using its cloud-based computational capabilities and extensive data access. The workflow included image collection, cloud masking, multi-scene compositing, boundary clipping, spatial resampling, and projection harmonization to ensure consistency and comparability across multi-temporal datasets. To reduce the influence of water bodies on RSEI calculation, we applied the Modified Normalized Difference Water Index (MNDWI) for water body masking, thereby enhancing the accuracy of ecological indicator classification.

2.2.2 Driving factors analysis

We derived the driving factors of the ecological environment from the TerraClimate dataset (Abatzoglou et al., 2018), which spans from 1958 to the present with a spatial resolution of approximately 4 km. TerraClimate was selected because it provides spatially continuous and consistent hydroclimatic variables that match the MODIS-based RSEI time series (2000–2023), reducing bias from multi-source data integration. The selected variables include annual average temperature (TAVG), Palmer Drought Severity Index (PDSI), Precipitation accumulation (PR), Downward surface shortwave radiation (SRAD), SOIL moisture (SOIL), Vapor pressure deficit (VPD), and Wind speed (VS).

Based on the GEE platform, monthly data for each variable were first acquired and composited over the growing season to better represent ecological conditions. Subsequently, the datasets were resampled to 1 km using the nearest-neighbor method, and values were extracted at the pixel locations of the RSEI grid to ensure parameter consistency and spatial comparability. Prior to model training, we systematically assessed multicollinearity among candidate climate variables. Specifically, we calculated pairwise Pearson correlation coefficients for all candidate climate variables and identified highly correlated pairs (|r| >0.85). For highly redundant pairs (e.g., VS and TAVG, SRAD and PR), we excluded one variable from each pair to mitigate multicollinearity effects. After this screening process, seven final indicators were retained for model input, ensuring low multicollinearity and improving model stability. A Pearson correlation heatmap is provided in Supplementary Notes S2 (Figure 4), which visualizes the relationships among the climate variables and supports the selection of the final indicators. Additional results can be found in Supplementary Notes S2.

Land use data were obtained from the annual China Land Cover Dataset (CLCD) developed by Wuhan University, which provides 30 m resolution products with an overall classification accuracy of 80%–85% (Yang and Huang, 2021). In this study, LUCC derived from CLCD was used as a proxy for human activity influences because it provides annually continuous, spatially consistent, and harmonized long-term information over 2000–2023.

2.3 Research methods

2.3.1 Ecological factor calculation

This study builds upon the RSEI model proposed by Xu Hanqiu and introduces several enhancements. To address saturation issues in regions with dense vegetation (Xu et al., 2019), the kNDVI was employed as an alternative to the conventional NDVI, enhancing the precise quantification of vegetation vigor (Wang et al., 2023a). The wetness parameter was computed through the Tasseled Cap Transformation (WET). At the same time, the aridity aspect was captured by the Normalized Difference Built-up and SOIL Index (NDBSI), and the temperature component was derived from the land surface temperature (LST) sourced from the MOD11A2 dataset (Wan et al., 2015).

kNDVI is calculated as shown in Equation 1.

$\text{kNDVI}=\tanh \left[{\left(\frac{{\mathrm{\rho }}_{\text{nir}1}-{\mathrm{\rho }}_{\text{red}}}{2\mathrm{\sigma }}\right)}^2\right](1)$

Where ${\mathrm{\rho }}_{\text{nir}1}$ denotes the reflectance in the near-infrared band, while ${\mathrm{\rho }}_{\text{red}}$ signifies the reflectance in the red band.

WET is calculated as shown in Equation 2.

$\text{wet}=0.1147{\mathrm{\rho }}_{\text{red}}+0.2489{\mathrm{\rho }}_{\text{blue}}+0.2408{\mathrm{\rho }}_{\text{green}}+0.3132{\mathrm{\rho }}_{\text{blue}}-0.3122{\mathrm{\rho }}_{\text{nir}2}-0.6416{\mathrm{\rho }}_{\text{swir}1}-0.508(2)$

Where ${\mathrm{\rho }}_{\text{red}}$, ${\mathrm{\rho }}_{\text{blue}}$, ${\mathrm{\rho }}_{\text{green}}$, ${\mathrm{\rho }}_{\text{nir}2}$, ${\mathrm{\rho }}_{\text{swir}1}$ and ${\mathrm{\rho }}_{\text{swir}2}$ represent the reflectance values of the red band, near-infrared band 1, blue band, green band, near-infrared band 2, shortwave infrared band 1, and shortwave infrared band 2, respectively.

NDBSI is calculated using Equations 3–5.

$\text{NDBSI}=\left(\text{IBI}+\text{SI}\right)/2(3)$

$\text{IBI}=\frac{2{\mathrm{\rho }}_{\text{swir}1}/\left({\mathrm{\rho }}_{\text{swir}1}+{\mathrm{\rho }}_{\text{nir}1}\right)-\left[{\mathrm{\rho }}_{\text{nir}1}/\left({\mathrm{\rho }}_{\text{nir}1}+{\mathrm{\rho }}_{\text{red}}\right)+{\mathrm{\rho }}_{\text{green}}/\left({\mathrm{\rho }}_{\text{green}}+{\mathrm{\rho }}_{\text{swir}1}\right)\right]}{2{\mathrm{\rho }}_{\text{swir}1}/\left({\mathrm{\rho }}_{\text{swir}1}+{\mathrm{\rho }}_{\text{nir}1}\right)+\left[{\mathrm{\rho }}_{\text{nir}1}/\left({\mathrm{\rho }}_{\text{nir}1}+{\mathrm{\rho }}_{\text{red}}\right)+{\mathrm{\rho }}_{\text{green}}/\left({\mathrm{\rho }}_{\text{green}}+{\mathrm{\rho }}_{\text{swir}1}\right)\right]}(4)$

$\text{SI}=\frac{\left[\left({\mathrm{\rho }}_{\text{swir}1}+{\mathrm{\rho }}_{\text{red}}\right)-\left({\mathrm{\rho }}_{\text{nir}1}+{\mathrm{\rho }}_{\text{blue}}\right)\right]}{\left[\left({\mathrm{\rho }}_{\text{swir}1}+{\mathrm{\rho }}_{\text{red}}\right)+\left({\mathrm{\rho }}_{\text{nir}1}+{\mathrm{\rho }}_{\text{blue}}\right)\right]}(5)$

where SI denotes the SOIL index, IBI denotes the building index, and ${\mathrm{\rho }}_{\text{blue}}$, ${\mathrm{\rho }}_{\text{green}}$, ${\mathrm{\rho }}_{\text{red}}$, ${\mathrm{\rho }}_{\text{nir}2}$, and ${\mathrm{\rho }}_{\text{swir}2}$ represent the reflectance values of the corresponding bands.

2.3.2 Construction of RSEI based on the entropy weight method

To improve the objectivity and sensitivity of the RSEI model, this study incorporates the entropy weight method to allocate weights to the standard four ecological indicators, thereby enhancing the rationality of the comprehensive ecological quality evaluation. The model comprises four ecological indicators: kNDVI, WET, NDBSI, and LST.

We first normalized all indicators to ensure comparability across different magnitudes and units. For positive indicators (greenness and wetness), min–max normalization was applied as follows:

${\mathrm{N}}_{\mathrm{i}}=\left({\mathrm{A}}_{\mathrm{i}}-{\mathrm{A}}_{\min }\right)/\left({\mathrm{A}}_{\max }-{\mathrm{A}}_{\min }\right)(6)$

${\mathrm{N}}_{\mathrm{i}}=\left({\mathrm{A}}_{\max }-{\mathrm{A}}_{\mathrm{i}}\right)/\left({\mathrm{A}}_{\max }-{\mathrm{A}}_{\min }\right)(7)$

Where Ni denotes the normalized value of the indicator, A represents the original value of the indicator at pixel i, and Amin and Amax are the minimum and maximum values of the indicator, respectively. Greenness and wetness were normalized using Equation 6, whereas dryness and heat were normalized using Equation 7.

These normalized indicators were then used as inputs for entropy-based weighting in the RSEI model. Weight Calculation Based on Information Entropy. Based on information entropy theory, we employed the entropy weight method to calculate the weight coefficients of each ecological indicator in the RSEI model. Information entropy measures the uncertainty associated with the probability of a random event. For a discrete random variable X with n possible states {X₁,X₂ ,…, X_n-1, X_n} and corresponding occurrence probabilities Pi, the entropy-based weighting procedure follows Equations 8–12.

$\mathrm{H}\left(X\right)=\mathrm{H}\left({\mathrm{P}}_1,{\mathrm{P}}_2,\cdots {\mathrm{P}}_{\mathrm{n}}\right)=-\mathrm{K}{\displaystyle \sum _{\mathrm{i}=1}^{\mathrm{m}}}{\mathrm{p}}_{\text{ij}}·\text{In}{\mathrm{p}}_{\text{ij}}(8)$

The information entropy, for the indicator, is defined as:

$\mathrm{K}=1/\text{Inm}(9)$

${\mathrm{P}}_{\text{wu}}={\mathrm{X}}_{\text{wu}}/{\displaystyle \sum _{\mathrm{u}=1}^{\mathrm{m}}}{\mathrm{X}}_{\text{wu}}(10)$

${\mathrm{e}}_{\mathrm{u}}=-\mathrm{k}{\displaystyle \sum _{\mathrm{u}=1}^{\mathrm{m}}}{\mathrm{P}}_{\text{wu}}·\ln {\mathrm{P}}_{\text{wu}}(11)$

Where K is the normalization coefficient, m denotes the total number of pixels, X_u is the indicator’s value at pixel u, X_v is the indicator’s value at pixel v, and Puv represents the probability of X_uv in the _u indicator.

The final weight W_u of the _u indicator is calculated based on its entropy as follows:

${\mathrm{w}}_{\mathrm{u}}=\left(1-{\mathrm{e}}_{\mathrm{u}}\right)/{\displaystyle \sum _{\mathrm{u}=1}^{\mathrm{n}}}\left(1-{\mathrm{e}}_{\mathrm{u}}\right)(12)$

We then constructed the RSEI by computing a weighted average of the normalized indicators using the EWM-derived weights. RSEI was computed as a weighted sum of normalized indicators (Equation 13). The computation formula is expressed as follows:

$\text{RSEI}={\displaystyle \sum _{\mathrm{i}=1}^{\mathrm{n}}}{\mathrm{w}}_{\mathrm{j}}·{\mathrm{N}}_{\mathrm{i}}(13)$

In the formula, n represents the number of indicators; N_i denotes the normalized value of each indicator; and w_j denotes the weight of each indicator. The final RSEI values are normalized to the range of [0, 1] and classified into five levels according to the Technical Specification for Ecological Environment Assessment (DB11/T 1877-2021): excellent (0.8–1), good (0.6–0.8), moderate (0.4–0.6), fair (0.2–0.4), and poor (0–0.2). To improve the stability of indicator weights in RSEI, this study determines weights using EWM. PCA, which is commonly used in conventional RSEI construction, is included only for robustness comparison and does not affect the final weights. A theoretical comparison and additional analyses are provided in the Supplementary Notes (Supplementary Section S1; Supplementary Figures S1–S3).

2.3.3 Stability analysis

To assess interannual variability and ecological stability across the study area, we calculated the pixel-level Coefficient of Variation (CV) of the RSEI. CV is defined as shown in Equation 14.

$\text{CV}=\frac{1}{\mathrm{X}}\sqrt{\frac{1}{\left(\mathrm{n}-1\right)}{\displaystyle \sum _{\mathrm{i}=1}^{\mathrm{n}}}{\left({\mathrm{X}}_{\mathrm{I}}-\overline{\mathrm{X}}\right)}^2}(14)$

Where $\overline{X}$ denotes the mean RSEI value over multiple years, n represents the total number of years in the time series. X_i is the RSEI value of the Yellow River Basin in year i. A lower CV indicates a minor interannual fluctuation and higher ecological stability, whereas a higher CV suggests more pronounced variability and lower stability in the ecosystem.

2.3.4 Trend analysis

This study employed a combination of the Sen–MK non-parametric test to quantitatively assess the temporal evolution of ecological quality in the Shaanxi section of the Yellow River Basin from 2000 to 2023 for trend analysis of the RSEI (Fernandes and Leblanc, 2005).

The Theil–Sen estimator is a non-parametric method suitable for trend detection in non-normally distributed remote sensing time series. It calculates the median of the slopes between all possible pairs of years to represent the overall trend. The Theil–Sen slope is calculated as shown in Equation 15.

$\mathrm{\beta }=\text{median}\left(\frac{{\text{RSEI}}_{\mathrm{j}}-{\text{RSEI}}_{\mathrm{i}}}{\mathrm{j}-\mathrm{i}}\right),\forall \mathrm{j}>\mathrm{i}(15)$

RSEI_i and RSEI_j are the RSE_i values in years i and j, respectively, and β denotes the overall trend slope.

The statistical significance of Theil-sen trend analysis is defined by MK’s Z value. When the absolute value of Z is greater than 1.65, 1.96, and 2.58, it means that the trend has passed the significance test with 90%, 95%, and 99% confidence, respectively. For the time series RSEI, the calculation method of standardized test statistic. The MK test statistics were calculated following Equations 16–19.

$\mathrm{Z}=\left\{\begin{array}{c}\frac{S-1}{\sqrt{\text{Var}\left(\mathrm{S}\right)}},S>0\\ 0,S=0\\ \frac{S+1}{\sqrt{\text{Var}\left(S\right)}},S<0\end{array}\right.(16)$

$S={\displaystyle \sum _{\mathrm{i}=1}^{\mathrm{n}-1}}{\displaystyle \sum _{\mathrm{j}=\mathrm{i}+1}^{\mathrm{n}}}\text{sign}\left({\text{RSEi}}_{\mathrm{J}}-{\text{RSEI}}_{\mathrm{i}}\right)(17)$

$\text{sign}\left(\mathrm{\theta }\right)=\left\{\begin{array}{c}1,\mathrm{\theta }>0\\ 0,\mathrm{\theta }=0\\ -1,\mathrm{\theta }<0\end{array}\right.(18)$

$\text{Var}\left(S\right)=\frac{\mathrm{n}\left(\mathrm{n}-1\right)\left(2\mathrm{n}+5\right)}{18}(19)$

Building on previous studies on the response mechanisms of NDVI to climate change, this study employs partial correlation coefficients and multiple correlation coefficients to explore the environmental driving mechanisms of the spatiotemporal evolution of RSEI. The statistical significance level for both partial and multiple correlation coefficients was set at 0.05 (|Z| > 1.96) (Lin et al., 2020).

2.3.5 Analysis of persistence and future trend prediction

The Hurst exponent, a statistical metric assessing the long-range correlation within a time series, has found extensive utility in sustainability research and the analysis of predictive patterns in ecological systems (Chen et al., 2025). Here, we utilized the Rescaled Range (R/S) analysis technique to calculate the Hurst exponent for the RSEI time series spanning from 2000 to 2023 at the pixel level. The Hurst exponent, ranging from 0 to 1, characterizes the behavior of a system. A Hurst exponent value below 0.5 indicates anti-persistent behavior, suggesting a propensity for future trends to reverse the current state. A Hurst exponent of 0.5 signifies a random walk, indicating a lack of sustained trend and unpredictability within the system. Conversely, a Hurst exponent exceeding 0.5 indicates persistent behavior, suggesting the likelihood of the system maintaining its current trend in the short term. A joint analysis of the Hurst exponent and Theil–Sen slope was conducted to determine the direction and stability of ecological evolution. The Theil–Sen slope signifies the historical trend, while the Hurst exponent indicates the sustainability of this trend. In this study, the ‘future’ indicated by the Hurst exponent refers to the near-term tendency after 2023, rather than a year-specific forecast. By integrating both metrics, ecological trends can be classified into four distinct categories: persistent improvement (Sen >0 and H > 0.5), continuous degradation (Sen <0 and H >0.5), potential reversal from improvement (Sen >0 but H <0.5), and potential reversal from degradation (Sen <0 and H <0.5). This classification system enhances our understanding of short-term ecological trajectories and serves as a foundation for evaluating regional ecosystem stability.

2.3.6 Evaluation index of model effect

To evaluate the predictive accuracy of each model, three metrics were employed: the coefficient of determination (R²), root mean square error (RMSE), and mean absolute error (MAE). R² assesses the model’s goodness of fit, with values ranging from ($-\infty$, 1]. A value closer to 1 signifies a stronger explanatory capability of the model’s variables for y, indicating a better alignment with the data. RMSE, the square root of the mean square error, reflects the standard deviation of the differences between predicted and actual values. On the other hand, MAE quantifies the average absolute difference between predicted and observed values, where smaller values denote enhanced model performance. This metric offers a more practical evaluation of predictive precision. Model performance metrics were calculated using Equations 20–22.

$R^2=1-\frac{{\displaystyle {\sum }_{i=1}^n}{\left(y_i-{\widehat{y}}_i\right)}^2}{{\displaystyle {\sum }_{i=1}^n}{\left(y_i-\overline{y}\right)}^2}(20)$

$RMSE=\sqrt{MSE}=\sqrt{\frac{1}{n}{\displaystyle \sum _{i=1}^n}{\left(y_i-{\widehat{y}}_i\right)}^2}(21)$

$MAE=\frac{1}{n}{\displaystyle \sum _{i=1}^n}\left|y_i-{\widehat{y}}_i\right|(22)$

where n represents the number of samples, ${\mathrm{y}}_{\mathrm{i}}$ represents the true value of the model, and ${\widehat{\mathrm{y}}}_{\mathrm{i}}$ represents the predicted value of the model.

2.3.7 Land use transfer matrix analysis

The land use transition matrix is essential for quantitatively characterizing the mutual conversion relationships among land use types over time. It serves to reveal the spatiotemporal dynamics of regional land use patterns. This approach assumes that land use change depends only on the initial and final states. By constructing a two-dimensional transition matrix, it systematically reflects the inflow and outflow, conversion direction, and transformation intensity of various land use categories within a specific period and can be formulated under a Markov-chain framework.

In this study, we used the annual CLCD released by Wuhan University (30 m resolution, overall classification accuracy 80%–85%) and selected the data for 2000 and 2023. The land cover classes were aggregated into seven major categories: grassland, cropland, forest, bare land, impervious surface, shrubland, and water, to construct the land use transition matrix.

After harmonizing the projection and spatial resolution and clipping the maps to the study area, we overlaid the 2000 and 2023 rasters and performed pixel-wise cross-tabulation to obtain the transition matrix $A=\left[A_{ij}\right]$, where Aij denotes the area converted from class i (in 2000) to class j (in 2023), with i,j = 1,2,3 …n and n representing the number of land use categories. Pixel counts were further converted to area (km²) using the pixel area (30 m × 30 m). By comparing each land use type’s transition paths and magnitude, the study aims to identify the driving mechanisms behind land use pattern changes, such as ecological restoration projects and urban expansion, thereby providing supporting evidence for interpreting the land-surface basis of spatiotemporal changes in RSEI.

2.3.8 Machine learning modeling and identification of driving factors

To further identify the dominant factors affecting RSEI variations and determine their importance ranking, we employed a machine-learning-based driver identification framework and compared three commonly used regression algorithms to select the most suitable model. To enhance reproducibility, the full workflow (data split, evaluation metrics, SHAP configuration) is explicitly described below, while detailed parameter settings are provided in the Supplementary Notes. To optimize each algorithm’s predictive performance, their parameters must be appropriately configured.

SVR built upon the framework of Support Vector Machines, seeks the optimal fitting hyperplane in high-dimensional feature spaces and exhibits strong generalization ability, making it suitable for high-dimensional, small-sample problems. XGBoost, an advanced member of the Boosting family, incorporates second-order gradient optimization and regularization mechanisms, enabling high precision and robustness in modeling complex nonlinear relationships. Based on the Bagging strategy, the RFR integrates multiple decision trees to enhance adaptability to multivariate nonlinear problems and ensures stable and reliable predictive performance (Breiman, 2001).

We used 1,400 samples with TAVG, PDSI, PR, SRAD, SOIL, VPD, and VS as predictors and RSEI as the target, split 80:20 for training and testing. Specifically, samples were randomly divided into training (80%) and testing (20%) subsets, and a fixed random seed (random_state = 42) was used to ensure reproducibility. XGBoost, Random Forest, and SVR were trained, and the best model was used for SHAP analysis. Model performance was evaluated on the testing set using R², RMSE, and MAE, and the model with the highest R² (with RMSE and MAE as complementary criteria) was selected for subsequent interpretation.

To enhance model interpretability, we incorporated the SHAP algorithm based on the Shapley value concept from cooperative game theory. SHAP enables the quantification of the marginal contribution of each input variable to the model’s prediction outcomes at both global and local scales (Jiang et al., 2024). It further reveals the response thresholds, influence magnitudes, and synergistic regulatory relationships among key driving factors, thus providing a robust foundation for causal inference of RSEI variation (Airiken and Li, 2024; Nagy and Molontay, 2024). For SHAP implementation, we used the Python shap package. Tree-based models (RFR and XGBoost) were interpreted using TreeExplainer, whereas SVR was interpreted using KernelExplainer with a randomly sampled background dataset (n = 100) to reduce computational burden. We used the Python shap package to visualize model outputs, enabling the identification of dominant variables, their optimal response intervals, and feature interaction effects.

To comprehensively evaluate the predictive performance of different models, this study adopts the R², RMSE, and MAE as quantitative metrics. The results indicate that the Random Forest Regression (RFR) significantly outperforms Support Vector Regression (SVR) and Extreme Gradient Boosting (XGBoost) in terms of fitting accuracy, achieving the highest R² value (0.9063), the lowest RMSE value (0.0609), and MAE results consistent with the R²trend, thereby demonstrating superior regional adaptability and modeling robustness. Consequently, the RFR was selected as the optimal model in this study. The SHAP method was subsequently employed to quantify the marginal contributions of each driving factor to the model output at both global and local scales, thereby identifying key drivers, response intensities, and threshold intervals to support the interpretation of dominant mechanisms underlying regional ecological variation (Airiken and Li, 2024; Nagy and Molontay, 2024). The performance comparison results of the three models are presented in Table 1. Detailed hyperparameter settings and the complete implementation procedure are provided in the Supplementary Notes to facilitate reproducibility.

Table 1

Table 1. Performance metrics of the SVR, XGBoost, and RFR.

3 Result

3.1 Interannual variation characteristics of RSEI in the Shaanxi section of the Yellow River basin

The annual average changes in the RSEI and its components in the study area (Figure 2) reveal that from 2000 to 2023, the RSEI averaged 0.51. The lowest recorded value was 0.45 in 2016, and the highest was 0.55 in 2021. The Theil–Sen analysis indicates a slight upward trend, with a mean slope of +0.0035 a⁻¹ (≈0.69% a⁻¹, relative to the multi-year mean RSEI of 0.51), indicating an overall improvement in ecological quality over this period.

Figure 2

Figure 2. Interannual trends of the optimized RSEI and its four constituent factors (kNDVI, WET, NDBSI, LST) in the Shaanxi section of the Yellow River Basin from 2000 to 2023.

The mean values indicate positive humidity and greenness indices, contrasted by negative dryness and heat indices. Notably, kNDVI and WET showed significant increases at rates of +0.0053/a and +0.0007/a, respectively, with kNDVI peaking at 0.37 in 2021. Conversely, LST and NDBSI declined gradually at rates of −0.0002/a and −0.0027/a, respectively.

The RSEI of the Shaanxi section of the Yellow River Basin has exhibited an overall increasing trend from 2000 to 2023 (Figure 3). Regionally, RSEI increased significantly in both southern and northern Shaanxi (p < 0.05), whereas the Guanzhong region showed a weak, non-significant trend (p > 0.05). Conversely, the Guanzhong region experienced an average annual growth rate of −0.49%, and the increasing trend did not reach statistical significance (p > 0.05). Significant differences in the spatial and temporal patterns of RSEI derived from EWM and PCA were observed (Supplementary Figures S1–S3). The overall agreement was low (overall correlation = 0.195; overall Kappa = 0.111), highlighting the superior temporal stability and classification consistency of the EWM approach (see Supplementary Notes).

Figure 3

Figure 3. Ecological quality changes in different regions.

3.2 Spatial distribution of RSEI in the Shaanxi section of the Yellow River basin

This study employs established research classification criteriato categorize ecological quality into five levels (poor, fair, moderate, good, excellent) with a grade interval of 0.2 (Yuan et al., 2021; Xia et al., 2025). The spatial distribution map illustrating ecological quality across different years is presented in Figure 4. Analysis of Figure 4 reveals a consistent pattern of lower remote sensing ecological indices in northern regions, predominantly falling within the extremely poor and poor categories. The Guanzhong region demonstrates a medium ecological index grade, while the southern Shaanxi region exhibits predominantly good and excellent grades. A discernible trend of decreasing ecological remote sensing indices with increasing latitude is observed. Notably, within the northern Guanzhong area, pockets of good-grade areas are identified on both the eastern and western peripheries, notably in the Ziwuling Forest Park to the west and the Huanglong Mountain Scenic Spot to the east.

Figure 4

Figure 4. Distribution map of ecological quality (a) RSEI values in 2000; (b) RSEI values in 2005; (c) RSEI values in 2010; (d) RSEI values in 2015; (e) RSEI values in 2020; (f) RSEI values in 2023.

In 2015, the region with a poor rating constituted 13.5% of the total area, primarily concentrated in the Maowusu Desert area in northwest Shaanxi Province. Subsequent time points showed minimal instances of poor ratings. Regions with poor ratings exhibited a declining trend, reaching their lowest point in 2023. Areas classified as medium encompassed approximately 35% of the study area, displaying a steady increase over time. The proportion of areas classified as good has been consistently rising. The percentage of regions rated as excellent was comparable to those rated as very poor initially but has since shown an increase.

3.3 Changes in RSEI in the Shaanxi section of the Yellow River basin

The distribution map of the multi-year average value of RSEI from 2000 to 2023 is shown in Figure 5a. The results indicate that the RSEI distribution in the study area exhibits significant spatial differences, generally presenting a spatial pattern of high values in the south and low values in the north. Areas with good ecological environment conditions are mainly distributed near the southern Shaanxi region and the Guanzhong Plain. In contrast, the ecological environment in the northern Shaanxi Loess Plateau and densely populated areas is relatively poor. The proportion of areas with an annual average RSEI greater than 0.6 reaches 70.20%, indicating that the ecological quality in most areas of the Shaanxi section of the Yellow River Basin is at a good level.

Figure 5

Figure 5. Spatial distribution, trend, and stability of the RSEI in the Shaanxi section of the Yellow River Basin during 2000–2023. (a) Spatial distribution of multi-year average RSEI values; (b) Spatial distribution of Sen’s slope estimates for RSEI trends; (c) Spatial distribution of Theil–Sen slope values indicating long-term RSEI trends; (d) CV of RSEI values indicating temporal stability.

To examine the ecological quality trend in the Shaanxi section of the Yellow River Basin from 2000 to 2023, this study utilized the Theil–Sen median slope method to estimate the rate of change based on pixel-scale RSEI data. The statistical significance of these estimates was assessed using the MK test. Table 2 outlines the statistical principles of the Sen–MK method. The linear change rate of RSEI values in the study area varied from −0.022 to 0.029 a⁻¹, with an overall increasing trend (Figure 5). The spatial distribution of Sen's slope is shown in Figure 5b. The average change rate was calculated at +0.0035 a⁻¹, indicating a slight overall improvement in regional ecological quality.

Table 2

Table 2. The statistical significance levels for the change trend of RSEI from 2000 to 2023 and the area proportion of each level.

The combined analysis of the Sen–MK results (Figure 5c) reveals that, at a significance level of p < 0.05, the changing trends of RSEI in the study area can be categorized into five groups:increase (I), significant increase (SI), decrease (D), significant decrease (SD), and stable (S). The I category is predominant, constituting 45.05% of the total, primarily concentrated in the central and southern Loess Plateau of northern Shaanxi and the Guanzhong Plain, including southern Yan’an, northern Xianyang, and the surrounding areas of Xi’an. The SI category represents 11.17% of the total and is predominantly distributed in Baoji, Weinan, and other regions within the Guanzhong Plain, indicating notable ecological restoration achievements in these areas. The D category accounts for 36.84%, while the SD category represents 6.20%, mainly distributed in the mountainous areas of central and western Shangluo and the hilly areas in the southern part of Weinan, where ecosystems face certain degradation pressures. The S category has a relatively low representation at 0.74% and is scattered across the area. Overall, these results suggest a gradual improvement in ecological condition in the Shaanxi section of the Yellow River Basin over the past 2 decades. However, targeted measures are still needed to address localized degradation and strengthen management in specific areas. The geographic locations of place names mentioned here are shown in Figure 1.

3.4 Stability analysis of RSEI in the Shaanxi section of the Yellow River basin

To evaluate the interannual fluctuation characteristics of the ecosystem, this study calculated the CV based on the RSEI time series from 2000 to 2023, thereby quantifying the stability level of ecological quality (Figure 5d). The results show that the minimum CV value of the RSEI in the study area is 0.004, the maximum value is 1.632, and the multi - year average value is 0.216. Overall, it presents a spatial pattern of “high in the north, low in the south, and fluctuating in the middle”.

Regarding spatial proportions, 44.28% of the surveyed area demonstrated CV values exceeding 0.20, indicating pronounced ecological variability. These regions were predominantly concentrated in the majority of Yulin, particularly in its northern and central sectors, the gully regions of northern and central Yan’an, and the southern mountainous areas of Weinan. Significant natural disturbances and human-induced pressures characterize these areas. Approximately 39.88% of the area exhibited CV values ranging from 0.10 to 0.20, signifying moderate fluctuation levels. These areas were mainly distributed across the northern periphery of the Guanzhong Plain, central Baoji, and the hilly terrains of Tongchuan. Merely 15.83% of the area displayed CV values below 0.10, indicating low interannual variability and high ecosystem stability. These stable regions were predominantly situated in the Qinba Mountains of southern Shaanxi and the ecologically pristine zones of southern Shangluo. The findings suggest that the ecological stability in the Shaanxi region of the Yellow River Basin is generally robust, with localized risks of ecological fluctuations observed in specific areas of the northern Shaanxi Plateau and the Guanzhong Plain.

3.5 Sustainability analysis of RSEI in the Yellow River basin

To further evaluate the long-term trends and sustainability of ecological quality, we calculated the Hurst exponent of the RSEI using the Rescaled Range (R/S) analysis method (Figure 6a). The findings indicate that the average Hurst index for each pixel in the Shaanxi section of the Yellow River Basin is 0.61, suggesting persistent behavior and that recent ecological changes are likely to continue in the near term. Notably, 91% of the area exhibits a Hurst index greater than 0.5, indicating persistence, with future changes likely mirroring the past 24 years. This region predominantly includes the Guanzhong Plain, southern basins of Shaanxi, and northern Shaanxi. Conversely, 9% of the area has a Hurst index below 0.5, reflecting anti-persistence, where future trends are expected to diverge from historical patterns. This analysis evaluates whether the 2000–2023 trend is likely to persist or reverse in the near term after 2023, rather than providing year-by-year forecasts.

Figure 6

Figure 6. Ecological sustainability assessment and future trend prediction based on the Hurst exponent. (a) Spatial distribution of the Hurst exponent; (b) Classification map of predicted ecological trends. It indicates near-term persistence or reversal after 2023.

The Hurst exponent was combined with Sen’s slope to classify future ecological trajectories (Figure 6b). Areas characterized as potential reversal from improvement (Sen >0, H <0.5) constitute the largest proportion (51.97%) and span much of the region. In contrast, potential reversal from degradation (Sen <0, H <0.5) accounts for 39.06% and is primarily concentrated in the Qinling–Bashan Mountains. Regions showing persistent improvement (Sen >0, H >0.5) represent 4.67% and are mainly located in the northwest of Yulin, the middle-east of Yan’an, and the transition zones among the Guanzhong Plain, Loess Plateau, and Qinling–Bashan Mountains. Meanwhile, continuous degradation (Sen <0, H >0.5) covers 4.30%, mainly distributed in northern Yan’an and the Guanzhong Plain, indicating areas where ecological deterioration may persist and thus warrant strengthened protection and intervention.

3.6 Driving factor analysis

Building upon identifying the spatiotemporal evolution patterns of ecological quality, this section aims to elucidate the underlying mechanisms further. Specifically, it explores the dominant factors, response characteristics, and interaction effects that have influenced changes in ecological quality in the Shaanxi section of the Yellow River Basin from 2000 to 2023, from two dimensions: land use pattern changes and natural environmental drivers.

3.6.1 Land use change characteristics and driving effects

From 2000 to 2023, the study area significantly changed land use patterns (Figures 7a,b). Spatially, cropland area experienced a notable decline, mainly in the Loess Plateau of northern Shaanxi and the Qinba mountainous region in the south. In contrast, forest cover expanded substantially, with newly afforested areas concentrated in the hilly zones of the northern plateau and the Qinling Mountains, indicating the effectiveness of ecological programs such as reforestation and natural forest protection. The area of impervious surfaces increased significantly, with the Guanzhong urban agglomeration, which is centered around Xi’an, emerging as a high-intensity development zone. Localized urban construction clusters have also formed in cities such as Yan’an and Yulin. Water bodies showed a slight decline in area, primarily due to urban expansion. We found that bare land was mainly distributed along the northwestern edge of Yulin in the Mu Us Desert, which exhibited an overall decreasing trend.

Figure 7

Figure 7. Spatial distribution of land use types in the study area from 2000 to 2023 (a) Land use distribution in 2000; (b) Land use distribution in 2023.

Regarding land structure changes, the dominant land use types in the study area in 2000 were grassland, cropland, and forest, accounting for 37.45%, 32.79%, and 25.70% of the total area, respectively. By 2023, the proportions of grassland and cropland had declined to 35.46% and 30.16%, while forest area increased to 30.26%, indicating a growth in ecologically restorative land use. The proportion of impervious surfaces rose from 1.56% to 3.49%, more than doubling, while the area of bare land significantly decreased from 1.76% to 0.13%. The proportions of shrubland, water bodies, and other land types fluctuated slightly, showing relatively minor overall changes.

Table 3 reveals that grassland and cropland areas decreased by 5.24% and 7.96% (relative to 2000), respectively, while forest cover increased by 17.58% (relative to 2000). These land-cover changes are consistent with the long-term improvement of RSEI and the concurrent evolution of its component indicators (Table. 3; Figure 2). The shift from cultivated land and grassland to forests has enhanced vegetation greenness and moisture, as reflected by the overall increasing trends in kNDVI and WET during 2000–2023 (Figure 2). The relatively high kNDVI level around 2021 indicates strengthened vegetation conditions within this overall upward trajectory, supporting vegetation recovery as an important contributor to the RSEI increase. Concurrently, reductions in bare land and impervious surfaces have alleviated dryness and heat pressures, which is consistent with the decreasing tendencies in NDBSI and LST over the study period (Figure 2). Collectively, these land-use changes have improved ecological quality by regulating key indicators such as greenness, humidity, dryness, and heat (Miao et al., 2024; Wang et al., 2025).

Table 3

Table 3. Land use transition matrix of the study area from 2000 to 2023.

3.6.2 Model interpretation based on SHAP analysis

To further identify the dominant driving factors and their potential nonlinear response mechanisms in ecological quality evolution, this study employed the RFR combined with SHAP to conduct interpretability analysis.

Figure 8 presents the feature importance ranking of ecological variables, reflecting the overall contribution of each factor to the model output. Figure 9 shows the SHAP summary plot, where each point represents the marginal contribution of a feature for a single sample. The x-axis indicates the SHAP value, and the color denotes the magnitude of the feature value (red for high values, blue for low values), thereby revealing the response trends and nonlinear interaction mechanisms of the ecological variables.

Figure 8

Figure 8. Feature importance ranking.

Figure 9

Figure 9. SHAP summary plot.

The results indicate that SOIL was the most influential variable among all features. Its SHAP values exhibited a wide distribution, mainly in the positive range, suggesting that higher SOIL moisture values contributed significantly to higher RSEI values. PR, VS, and the PDSI followed as key meteorological drivers. In contrast, VPD, SRAD and TAVG showed lower contributions, with some exhibiting adverse effects on RSEI.

Figure 10 shows the SHAP interaction plots, illustrating the response relationships between dominant driving factors and RSEI. The results indicate that the SHAP values for SOIL increase markedly within the range of 20–40, suggesting a substantial positive contribution of moderate SOIL moisture to ecological quality. When interacting with VS, the SHAP values increase further as VS rises, indicating that appropriate wind levels can amplify the ecological regulatory effect of SOIL moisture. PR shows notably higher SHAP values within the 500–700 mm range, suggesting that annual precipitation within this optimal threshold improves regional ecological conditions. However, under high VPD conditions, SHAP values drop sharply, indicating that dry climates weaken the positive ecological effect of precipitation.

Figure 10

Figure 10. SHAP interaction plots illustrating the response relationships between dominant driving factors and RSEI.

When considering VS as the primary variable, its SHAP values consistently rise within the low-to-moderate range, indicating its ecological advantages. Nonetheless, in high VPD conditions, the SHAP values of VS decline, indicating that air dryness constrains its ecological impact.

Within the PDSI framework, SHAP values ranging from −2 to −1 demonstrate pronounced positive impacts, suggesting that moderate drought conditions may trigger beneficial adaptive reactions in vegetation. Specifically, in conjunction with wind speed, higher values of SHAP for PDSI correspond to an escalation in wind speed, underscoring the reinforcing effect of wind during moderate drought episodes. Conversely, under dry and windy circumstances, PDSI’s SHAP values may exhibit more substantial negative repercussions, indicating a compounded ecological strain.

SHAP values exhibit significant variability within the range of 2000–2200 W/m² for SRAD. When considering wind speed as the interacting variable, SHAP values demonstrate an upward trend with increasing VS, suggesting that moderate wind speeds can mitigate ecological stress caused by radiation and improve ecosystem regulation.

Elevated temperatures negatively impact the RSEI, indicating a potential decrease in ecological stability. Conversely, increased SOIL moisture levels are associated with higher SHAP values, suggesting that sufficient SOIL water availability may mitigate the detrimental impacts of higher temperatures.

SOIL moisture, precipitation, and wind speed are recognized as primary positive influencers within appropriate ranges. Nevertheless, their ecological impacts are influenced by temperature and atmospheric dryness, indicating ecosystems’ sensitivity and nonlinear responses to interactions among multiple factors.

4 Discussion

4.1 Spatiotemporal variations in RSEI

Shaanxi Province, a core ecological zone in the middle reaches of the Yellow River, encompasses three major climatic zones (warm temperate, temperate, and subtropica) and features diverse geomorphological types (Liu et al., 2020a; Li et al., 2024). Consequently, its ecosystems’ ecological carrying capacity and restoration potential demonstrate significant spatial heterogeneity. Northern Shaanxi experiences an arid climate and severe SOIL erosion, further challenged by intensive development of energy and chemical industries (Wang et al., 2021; Song et al., 2022). The Guanzhong Plain, characterized by its flat terrain, has undergone substantial urbanization and increased pollutant emissions during the past 2 decades. Southern Shaanxi, encompassing mountainous and valley regions, is subject to ecological degradation caused by mining activities due to its complex topography.

Following the implementation of ecological protection measures, research indicates that from 2000 to 2023, Shaanxi’s ecological quality has steadily improved, with its RSEI growth rate among the highest in China. Against this background, regional-scale dynamic ecological assessments are necessary to evaluate restoration outcomes and to support ecological protection policies in the Yellow River Basin. These assessments provide crucial evidence for advancing provincial ecological civilization and achieving the national “Great Protection of the Yellow River strategy.” Notably,by incorporating kNDVI and the EWM, this study improves the sensitivity and objectivity of RSEI in detecting ecological changes. These modifications prove particularly effective in regions with dense vegetation cover and complex terrain, thereby enhancing the responsiveness of the index to regional heterogeneity and providing a more reliable basis for trend interpretation and vulnerability identification.

This study developed an RSEI model using long-term, high-temporal-resolution MODIS data to analyze the spatiotemporal dynamics and sustainability of ecological quality in the Shaanxi section of the Yellow River Basin. Results revealed that the overall mean RSEI in the study area was 0.51, with a Theil–Sen slope of +0.0035 a⁻¹ (≈0.69% a⁻¹, relative to the multi-year mean RSEI of 0.51), demonstrating a pattern of early-stage fluctuation followed by late-stage recovery. Spatially, 66.75% of the area exhibited significant ecological improvement, primarily in Yan’an and Yulin in northern Shaanxi. In areas under reforestation and grassland restoration, the RSEI growth rate reached 0.84%/a, demonstrating the effectiveness of ecological policies on the Loess Plateau (Liu et al., 2020b; He et al., 2024). In addition, statistical tests reported in the Supplementary Notes indicate that EWM provides more stable weight allocation and better interannual consistency than PCA, which strengthens the reliability of multi-temporal ecological quality assessments.

Based on Sen–MK trend analysis and Hurst index estimation, 46.04% of the area may transition from improvement to degradation, with the total degraded area potentially reaching 52.62% (Li et al., 2025). The degradation hotspots identified in this study are spatially consistent with patches reported in previous studies across the Weihe River Basin, the Qinling Mountains, and northern Shaanxi mining areas, with similar temporal trends and driving mechanisms (Zhang et al., 2022; Liu et al., 2024a; Zhu et al., 2024). These findings indicate that ecosystems in Shaanxi remain vulnerable. Rapid urbanization and population growth intensify ecological stress, diminishing ecosystem stability and increasing vulnerability and degradation risk. Despite continued overall ecological improvement, significant spatial heterogeneity exists in ecosystem responses. Certain subregions demonstrate limited resilience to external disturbances and are susceptible to degradation under urban expansion and population concentration. Targeted spatial identification and differentiated management of vulnerable zones require immediate attention (Li et al., 2025). It is advisable to establish regional governance within the parameters of the Yellow River Protection Law and enact the following complementary measure. First, prioritize SOIL and water conservation in sloping farmland and loess hilly–gully areas, and optimize irrigation scheduling to reduce erosion and improve water-use efficiency. Second, in the grassland–cropland transition zone, restore shelterbelts, implement rotational grazing, and strengthen sand-control practices to stabilize the landscape. Third, in suburban and peri-urban areas, develop green corridors, deploy sponge-city facilities, and regulate impervious surface coverage to mitigate heat and runoff pressures. Finally, delineate riparian buffer zones and advance reclamation and ecological restoration in riverbanks and reclaimed mining areas to reduce localized degradation risks. These interventions are crucial for mitigating the risk of regression and fortifying the current positive trajectory of improvement.

4.2 Driving factors of RSEI

Mechanistically, ecological quality changes reflect the combined effects of hydroclimatic constraints and LUCC-driven land-surface alterations: climate regulates water–energy availability, while land-use transitions modify land-cover composition and surface properties, and the two jointly shape RSEI and its components. Accordingly, our results suggest that land-use pattern transformation and natural variables act synergistically to drive the spatiotemporal evolution of ecological quality. From 2000 to 2023, cropland and grassland areas decreased by 7.96% and 5.24%, respectively, while forest area increased substantially by 17.58%, predominantly in northern Shaanxi and Qinling regions (Shi et al., 2021). These patterns demonstrate the long-term effectiveness of ecological restoration projects such as reforestation and afforestation. Furthermore, the proportion of impervious surfaces increased from 1.56% to 3.49%, indicating significant urban construction land expansion. The conversion from cropland and grassland to urban land was notably concentrated, highlighting the growing anthropogenic influence on regional ecological landscape patterns (Albert et al., 2023; Chen et al., 2023).

Based on the results of the RFR combined with SHAP analysis, SOIL and PR emerged as the primary natural drivers influencing RSEI variation (Yuan et al., 2021). We found that both factors exerted positive ecological regulatory effects within their optimal thresholds: RSEI values increased markedly when annual precipitation ranged between 500 and 700 mm. Similarly, RSEI tended to be higher when SOIL remained within approximately 20%–40%, indicating that water availability is a key constraint on ecological restoration in this region. These findings suggest that water availability remains a critical constraint on regional ecological restoration.

Further analysis revealed significant interactions among the major ecological driving factors. Under moderate to low wind speed conditions, synergistic effects between VS and variables such as PR and SOIL were pronounced, contributing to enhanced regulatory capacity and ecological stability. However, when high VPD or persistently elevated temperatures occurred, the regulatory effects of these variables weakened substantially, reflecting complex nonlinear interaction mechanisms among ecological variables. Specifically, VS positively influenced ecological stability at moderate to low levels and amplified the ecological benefits of SOIL, PR, and the PDSI. Our analysis showed that high VPD or elevated temperatures significantly diminished the positive effects of wind speed and precipitation, indicating that hot and dry stress exerts a synergistic negative impact on the ecosystem. As the average TAVG increased, RSEI values tended to decline, suggesting that rising temperatures may exacerbate evapotranspiration processes and hinder ecological recovery in the region.

These findings not only clarify the mechanisms driving ecological change but also offer practical implications for future management. Based on the identified ecological hotspots and key driver responses, this study proposes brief governance recommendations. Priority should be given to ecological restoration through cropland-to-forest conversion and small watershed rehabilitation to improve land-use patterns; A monitoring and evaluation framework integrating RSEI and key drivers should be established to provide scientific support for basin-scale ecological restoration and sustainable development. A limitation is that LUCC was used as a proxy for socio-economic influences; future work could incorporate more direct socio-economic indicators (e.g., night-time lights, population density, and infrastructure intensity) when consistent long-term datasets are available for the full study period.

5 Conclusion

This study focuses on the Shaanxi section of the Yellow River Basin, utilizing long-term MODIS remote sensing data and the Google Earth Engine platform to construct an optimized RSEI that integrates the kNDVI and the entropy weight method. We systematically assessed the spatiotemporal patterns, stability, and sustainability of ecological quality. We employed the RFR and SHAP algorithm to identify the dominant ecological driving mechanisms. The main conclusions are as follows:

1. The annual average RSEI value in the study area exhibited a slightly fluctuating upward trend, increasing from 0.501 in 2000 to 0.515 in 2023. Over this period, the proportion of land area with extremely poor and poor ecological quality decreased, while the proportion of areas classified as medium, good, and excellent quality increased substantially. The significant decrease in extremely poor areas and increase in excellent areas suggest a notable improvement in the ecological quality of the Shaanxi section of the Yellow River Basin from 2000 to 2023.

2. The CV and Sen–MK analyses reveal that about 56.96% of the regions exhibit an improving trend, albeit with significant ecological volatility. Hurst exponent analysis further indicates that 39.2% of the improving-trend area (Sen >0) shows anti-persistence (H <0.5), implying that these improvements may be unstable and could reverse in the future.

3. The land use in the study area underwent substantial transformations between 2000 and 2023. Cultivated land decreased from 32.79% to 30.16%, while forest cover expanded from 25.70% to 30.26%. Impervious surfaces increased from 1.56% to 3.49%. The conversion of cultivated and grassland areas to forest enhanced greenness and humidity, contributing to an improvement in the RSEI. Bare land decreased from 1.76% to 0.13%, effectively mitigating dryness and heat, and overall ecological quality was enhanced.

4. The RFR–SHAP model analysis identifies SOIL moisture (20–40), annual precipitation (500–700 mm), and wind speed (1–3 m/s) as key natural drivers influencing RSEI changes. Under high temperature and VPD conditions, the regulatory impact of these positive factors diminishes, highlighting the nonlinear interactions among multiple factors.

Data availability statement

Publicly available datasets were analyzed in this study. This data can be found here: The data used in this study are publicly available from the following sources: MODIS remote sensing products were accessed via the Google Earth Engine platform (https://earthengine.google.com/); climatic variables were obtained from the TerraClimate dataset (https://www.climatologylab.org/terraclimate.html); and land use/land cover data were derived from the China Land Cover Dataset (CLCD), available at http://www.resdc.cn.

Author contributions

CM: Writing – original draft, Data curation, Software, Writing – review and editing. CZ: Writing – review and editing, Funding acquisition, Project administration, Data curation. HC: Data curation, Writing – original draft, Software, Investigation. MM: Software, Writing – review and editing, Writing – original draft, Data curation. SW: Data curation, Visualization, Investigation, Writing – original draft.

Funding

The author(s) declared that financial support was received for this work and/or its publication. This research was funded by Education Department of Shaanxi Provincial Government, grant number 24jk0336; The Key Research and Development Program of Xianyang, grant number L2024-ZDYF-ZDYF-GY-0036; Scientific Research Project of Shaanxi College of Communications Technology, grant number YJ24003.

Acknowledgements

We thank all the reviewers for their valuable comments and suggestions.

Conflict of interest

The author(s) declared that this work was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Generative AI statement

The author(s) declared that generative AI was not used in the creation of this manuscript.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Supplementary material

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fenvs.2026.1721306/full#supplementary-material

Footnotes

^{Abbreviations:}RSEI, Remote Sensing Ecological Index; kNDVI, Kernel-based Normalized Difference Vegetation; RFR, Random Forest Regression; SHAP, Shapley Additive Explanations; SOIL, SOIL moisture; PR, Precipitation accumulation; VS, Wind speed at 10 m; VPD, Vapor pressure deficit; TAVG, Average Temperature.

References

Abatzoglou, J. T., Dobrowski, S. Z., Parks, S. A., and Hegewisch, K. C. (2018). TerraClimate, a high-resolution global dataset of monthly climate and climatic water balance from 1958–2015. Sci. Data 5, 170191. doi:10.1038/sdata.2017.191

PubMed Abstract | CrossRef Full Text | Google Scholar

Airiken, M., and Li, S. (2024). The dynamic monitoring and driving forces analysis of ecological environment quality in the Tibetan Plateau based on the google earth engine. Remote Sens. 16, 682. doi:10.3390/rs16040682

CrossRef Full Text | Google Scholar

Albert, J. S., Carnaval, A. C., Flantua, S. G. A., Lohmann, L. G., Ribas, C. C., Riff, D., et al. (2023). Human impacts outpace natural processes in the amazon. Science 379, eabo5003. doi:10.1126/science.abo5003

PubMed Abstract | CrossRef Full Text | Google Scholar

Antonini, A. S., Tanzola, J., Asiain, L., Ferracutti, G. R., Castro, S. M., Bjerg, E. A., et al. (2024). Machine learning model interpretability using SHAP values: application to igneous rock classification task. Appl. Comput. Geosci. 23, 100178. doi:10.1016/j.acags.2024.100178

CrossRef Full Text | Google Scholar

Breiman, L. (2001). Random forests. Mach. Learn. 45, 5–32. doi:10.1023/A:1010933404324

CrossRef Full Text | Google Scholar

Cai, Z., Zhang, Z., Zhao, F., Guo, X., Zhao, J., Xu, Y., et al. (2023). Assessment of eco-environmental quality changes and spatial heterogeneity in the yellow river delta based on the remote sensing ecological index and geo-detector model. Ecol. Inf. 77, 102203. doi:10.1016/j.ecoinf.2023.102203

CrossRef Full Text | Google Scholar

Chen, S., Zhang, Q., Chen, Y., Zhou, H., Xiang, Y., Liu, Z., et al. (2023). Vegetation change and eco-environmental quality evaluation in the Loess plateau of China from 2000 to 2020. Remote Sens. 15, 424. doi:10.3390/rs15020424

CrossRef Full Text | Google Scholar

Chen, L., Chai, H., Du, W., Lian, Z., Wang, Y., Li, C., et al. (2025). Spatiotemporal dynamics of ecological quality and its drivers in shanxi province and its planned mining areas. Sci. Rep. 15, 31023. doi:10.1038/s41598-025-15550-3

PubMed Abstract | CrossRef Full Text | Google Scholar

Fang, C., Fan, Y., Bao, C., Li, G., Wang, Z., Sun, S., et al. (2023). China’s improving total environmental quality and environment-economy coordination since 2000: progress towards sustainable development goals. J. Clean. Prod. 387, 135915. doi:10.1016/j.jclepro.2023.135915

CrossRef Full Text | Google Scholar

Fernandes, R., and Leblanc, G. (2005). Parametric (modified least squares) and non-parametric (theil–sen) linear regressions for predicting biophysical parameters in the presence of measurement errors. Remote Sens. Environ. 95, 303–316. doi:10.1016/j.rse.2005.01.005

CrossRef Full Text | Google Scholar

He, Y., Ye, B., He, J., Wang, H., and Zhou, W. (2024). Long time-series monitoring and drivers of eco-quality in the upper-middle fen river basin of the eastern loess plateau: an analysis based on a remote sensing ecological index and google earth engine. Land 13, 2239. doi:10.3390/land13122239

CrossRef Full Text | Google Scholar

Hosseiny, B., Abdi, A. M., and Jamali, S. (2022). Urban land use and land cover classification with interpretable machine learning – a case study using sentinel-2 and auxiliary data. Remote Sens. Appl. Soc. Environ. 28, 100843. doi:10.1016/j.rsase.2022.100843

CrossRef Full Text | Google Scholar

Jiang, S., Sweet, L., Blougouras, G., Brenning, A., Li, W., Reichstein, M., et al. (2024). How interpretable machine learning can benefit process understanding in the geosciences. Earth’s Future 12, e2024EF004540. doi:10.1029/2024EF004540

CrossRef Full Text | Google Scholar

Li, J., He, S., Wu, H., Na, J., and Ding, H. (2024). Evaluation of geomorphological classification uncertainty using rough set theory: a case study of shaanxi province, China. Earth Surf. Process. Landforms 49, 4532–4548. doi:10.1002/esp.5965

CrossRef Full Text | Google Scholar

Li, J., Zhang, Y., Yang, L., and Shan, Z. (2025). Seasonal variations in ecological environment quality across different geomorphological regions and their response mechanisms to climate change. Sci. Rep. 15, 26385. doi:10.1038/s41598-025-11442-8

PubMed Abstract | CrossRef Full Text | Google Scholar

Liao, W., and Jiang, W. (2020). Evaluation of the spatiotemporal variations in the eco-environmental quality in China based on the remote sensing ecological index. Remote Sens. 12, 2462. doi:10.3390/rs12152462

CrossRef Full Text | Google Scholar

Lin, X., Niu, J., Berndtsson, R., Yu, X., Zhang, L., and Chen, X. (2020). NDVI dynamics and its response to climate change and reforestation in northern China. Remote Sens. 12, 4138. doi:10.3390/rs12244138

CrossRef Full Text | Google Scholar

Liu, D., Chen, H., Zhang, H., Geng, T., and Shi, Q. (2020a). Spatiotemporal evolution of landscape ecological risk based on geomorphological regionalization during 1980–2017: a case study of shaanxi province, China. Sustainability 12, 941. doi:10.3390/su12030941

CrossRef Full Text | Google Scholar

Liu, Z., Wang, J., Wang, X., and Wang, Y. (2020b). Understanding the impacts of ‘grain for green’ land management practice on land greening dynamics over the loess plateau of China. Land Use Policy 99, 105084. doi:10.1016/j.landusepol.2020.105084

CrossRef Full Text | Google Scholar

Liu, P., Ren, C., Yu, W., Ren, H., and Xia, C. (2023). Exploring the ecological quality and its drivers based on annual remote sensing ecological index and multisource data in northeast China. Ecol. Indic. 154, 110589. doi:10.1016/j.ecolind.2023.110589

CrossRef Full Text | Google Scholar

Liu, J., Xie, T., Lyu, D., Cui, L., and Liu, Q. (2024a). Analyzing the spatiotemporal dynamics and driving forces of ecological environment quality in the qinling mountains, China. Sustainability 16, 3251. doi:10.3390/su16083251

CrossRef Full Text | Google Scholar

Liu, Y., Xiang, W., Hu, P., Gao, P., and Zhang, A. (2024b). Evaluation of ecological environment quality using an improved remote sensing ecological index model. Remote Sens. 16, 3485. doi:10.3390/rs16183485

CrossRef Full Text | Google Scholar

Meng, X., Li, S., Akhmadi, K., He, P., and Dong, G. (2024). Trends, turning points, and driving forces of desertification in global arid land based on the segmental trend method and SHAP model. GISci. Remote Sens. 61, 2367806. doi:10.1080/15481603.2024.2367806

CrossRef Full Text | Google Scholar

Miao, W., Chen, Y., Kou, W., Lai, H., Sazal, A., Wang, J., et al. (2024). The HANTS-fitted RSEI constructed in the vegetation growing season reveals the spatiotemporal patterns of ecological quality. Sci. Rep. 14, 14686. doi:10.1038/s41598-024-65659-0

PubMed Abstract | CrossRef Full Text | Google Scholar

Nagy, M., and Molontay, R. (2024). Interpretable dropout prediction: towards XAI-based personalized intervention. Int. J. Artif. Intell. Educ. 34, 274–300. doi:10.1007/s40593-023-00331-8

CrossRef Full Text | Google Scholar

Qin, G., Wang, N., Wu, Y., Zhang, Z., Meng, Z., and Zhang, Y. (2024). Spatiotemporal variations in eco-environmental quality and responses to drought and human activities in the middle reaches of the Yellow river basin, China from 1990 to 2022. Ecol. Inf. 81, 102641. doi:10.1016/j.ecoinf.2024.102641

CrossRef Full Text | Google Scholar

Shi, S., Yu, J., Wang, F., Wang, P., Zhang, Y., and Jin, K. (2021). Quantitative contributions of climate change and human activities to vegetation changes over multiple time scales on the Loess plateau. Sci. Total Environ. 755, 142419. doi:10.1016/j.scitotenv.2020.142419

PubMed Abstract | CrossRef Full Text | Google Scholar

Song, Y., Xia, S., Xue, D., Luo, S., Zhang, L., and Wang, D. (2022). Land space change process and its eco-environmental effects in the guanzhong plain urban agglomeration of China. Land 11, 1547. doi:10.3390/land11091547

CrossRef Full Text | Google Scholar

Wan, Z., Hook, S., and Hulley, G. (2015). MOD11A2 MODIS/terra land surface temperature/emissivity 8-day L3 global 1km SIN grid V006. doi:10.5067/MODIS/MOD11A2.006

CrossRef Full Text | Google Scholar

Wang, K., Deng, L., Shangguan, Z., Chen, Y., and Lin, X. (2021). Sustainability of eco-environment in semi-arid regions: lessons from the Chinese Loess plateau. Environ. Sci. Policy 125, 126–134. doi:10.1016/j.envsci.2021.08.025

CrossRef Full Text | Google Scholar

Wang, Q., Moreno-Martínez, Á., Muñoz-Marí, J., Campos-Taberner, M., and Camps-Valls, G. (2023a). Estimation of vegetation traits with kernel NDVI. ISPRS J. Photogramm. Remote Sens. 195, 408–417. doi:10.1016/j.isprsjprs.2022.12.019

CrossRef Full Text | Google Scholar

Wang, X., Zhang, S., Zhao, X., Shi, S., and Xu, L. (2023b). Exploring the relationship between the eco-environmental quality and urbanization by utilizing sentinel and landsat data: a case study of the yellow river basin. Remote Sens. 15, 743. doi:10.3390/rs15030743

CrossRef Full Text | Google Scholar

Wang, Z., Yu, Y., Cheng, Y., Liu, D., Cheng, Y., and Pereira, P. (2025). Land use and RSEI spatial-temporal changes in horqin sandland (inner Mongolia, China). Ecol. Front. 45, 849–861. doi:10.1016/j.ecofro.2025.02.008

CrossRef Full Text | Google Scholar

Xia, J., Zhang, G., Ma, S., and Pan, Y. (2025). Spatial heterogeneity of driving factors in multi-vegetation indices RSEI based on the XGBoost-SHAP model: a case study of the jinsha river basin, yunnan. Land 14, 925. doi:10.3390/land14050925

CrossRef Full Text | Google Scholar

Xu, H., Wang, Y., Guan, H., Shi, T., and Hu, X. (2019). Detecting ecological changes with a remote sensing based ecological index (RSEI) produced time series and change vector analysis. Remote Sens. 11, 2345. doi:10.3390/rs11202345

CrossRef Full Text | Google Scholar

Yang, J., and Huang, X. (2021). The 30 m annual land cover dataset and its dynamics in China from 1990 to 2019. Earth Syst. Sci. Data 13, 3907–3925. doi:10.5194/essd-13-3907-2021

CrossRef Full Text | Google Scholar

Yang, X., Meng, F., Fu, P., Zhang, Y., and Liu, Y. (2021). Spatiotemporal change and driving factors of the eco-environment quality in the Yangtze river basin from 2001 to 2019. Ecol. Indic. 131, 108214. doi:10.1016/j.ecolind.2021.108214

CrossRef Full Text | Google Scholar

Yuan, B., Fu, L., Zou, Y., Zhang, S., Chen, X., Li, F., et al. (2021). Spatiotemporal change detection of ecological quality and the associated affecting factors in dongting lake basin, based on RSEI. J. Clean. Prod. 302, 126995. doi:10.1016/j.jclepro.2021.126995

CrossRef Full Text | Google Scholar

Zhang, K., Feng, R., Zhang, Z., Deng, C., Zhang, H., and Liu, K. (2022). Exploring the driving factors of remote sensing ecological index changes from the perspective of geospatial differentiation: a case study of the weihe river basin, China. Int. J. Environ. Res. Public Health 19, 10930. doi:10.3390/ijerph191710930

PubMed Abstract | CrossRef Full Text | Google Scholar

Zhou, M., Li, Z., Gao, M., Zhu, W., Zhang, S., Ma, J., et al. (2024). Revealing the eco-environmental quality of the yellow river basin: trends and drivers. Remote Sens. 16, 2018. doi:10.3390/rs16112018

CrossRef Full Text | Google Scholar

Zhou, Y., Cao, W., and Zhou, J. (2025). Dynamic assessment of eco-environmental quality in xiong’an new area, China using WB-RSEI new model. Sci. Rep. 15, 7592. doi:10.1038/s41598-025-91304-5

PubMed Abstract | CrossRef Full Text | Google Scholar

Zhu, Z., Cao, H., Yang, J., Shang, H., and Ma, J. (2024). Ecological environment quality assessment and spatial autocorrelation of northern shaanxi mining area in China based-on improved remote sensing ecological index. Front. Environ. Sci. 12, 1325516. doi:10.3389/fenvs.2024.1325516

CrossRef Full Text | Google Scholar