Soil salinization is a widespread form of land degradation in arid and semi-arid regions (Kumar et al., 2024). This phenomenon not only inhibits vegetation growth and accelerates land degradation, but also undermines desertification control efforts, triggers secondary degradation, and threatens agricultural production and infrastructure. (Yin et al., 2022; Sahbeni et al., 2023). Therefore, the degree of soil salinization should be incorporated as a key indicator in assessing the EEQ of IMWSA. Previous studies have shown that in semi-arid and arid regions, the Salinity Index-Transformed (SI-T) has the highest correlation with soil salt content and is more applicable to sparsely vegetated lands (Allbed et al., 2014; Wang et al., 2019). Therefore, this study incorporates SI-T as a salinity index into the classic RSEI model to more comprehensively characterize the impact of soil salinization on EEQ and enhance the sensitivity of RSEI to ecological changes in the study area. Table 2 outlines the equations for computing the various indicators (Ye et al., 2025; Zhang et al., 2024c).

The functional expression of the constructed eRSEI is given as Equation 1: e RSEI = f NDVI , WET , NDBSI , LST , SI − T (1)

Prior to index integration and PCA, normalization was performed to eliminate dimensional differences and to avoid weight imbalance among principal components caused by differing indicator units. The specific formula is given in Equation 2 (Zhang et al., 2022): N I i = I i − I min / I max − I min (2) where, NIi is the normalized value of the i th indicator, Ii is the original value of pixel i, and I _max and I _min are the maximum and minimum values of the indicator in the study area, respectively.

Based on the PCA, the five indicators mentioned above were integrated, and the calculation method for eRSEI was determined based on the sign of the factor loadings of the first principal component (PC1), as shown in Equation 3. e RSEI 0 = P C 1 f NDVI , WET , NDBSI , LST , SI − T   V NDVI > 0 , V W E T > 0 1 − P C 1 f NDVI , WET , NDBSI , LST , SI − T   V NDVI < 0 , V W E T < 0 (3) where, eRSEI₀ represents the initial eRSEI value, PC1 denotes the first principal component, and V _NDVI and V _WET correspond to the eigenvector values of NDVI and WET in PC1, respectively. To facilitate interannual comparisons, eRSEI values were normalized to the [0, 1] range, values near one suggest favorable EEQ, whereas values near 0 point to poorer conditions. Following previous studies (Xu et al., 2018), the normalized eRSEI was classified into five levels: (0.0–0.2], (0.2–0.4], (0.4–0.6], (0.6–0.8], and (0.8–1.0], corresponding to poor, fair, moderate, good, and excellent EEQ, respectively.

A combination of Theil-Sen Median and Mann-Kendall test methods was adopted to quantitatively analyze the spatiotemporal variation trends of eRSEI in IMWSA. The Sen slope estimation method characterizes the intensity of change trends by calculating the median slope of time series data, with its mathematical expression being shown in Equation 4 (Liu et al., 2025): β = median f j − f i / j − i , ∀ j > i (4) where f represents the observed values of the eRSEI time series, and median is the median statistical measure. When β > 0, it characterizes an upward trend in eRSEI, while negative values indicate a downward trend.

The Mann-Kendall non-parametric test method was used for significance testing, implemented through the standardized test statistic Z: when |Z| ≤Z_1-α/2, the null hypothesis is accepted, when |Z|>Z_1-α/2, the null hypothesis is rejected. This study sets the significance level at α = 0.05, corresponding to the two-sided test critical value Z = ±1.96. On this basis, the variation trend of eRSEI can be classified into five categories: significant increase, non-significant increase, no obvious change, non-significant decrease, and significant decrease. (Haas et al., 2025; Xu et al., 2025).

The coefficient of variation ( C v ) was used to evaluate the temporal variability of eRSEI in IMWSA, as shown in Equation 5. ( C v ) was used to evaluate the temporal variability of eRSEI in IMWSA as: C v = 1 x ¯ 1 n − 1 ∑ i = 1 n x i − x ¯ 2 (5) where C v represents the coefficient of variation, i is the time series, n is the total number of years, x i represents the eRSEI of the ith year, and x ¯ represents the average eRSEI value for all years within this period. The larger the C v , the stronger the fluctuation of eRSEI. C v is classified into five levels: low variation ( C v ≤0.10), relatively low variation (0.10< C v ≤0.20), moderate variation (0.20< C v ≤0.30), relatively strong variation (0.30< C v ≤0.40), and strong variation ( C v >0.40).

The Hurst exponent (H) was calculated based on the rescaled range method (R/S method), which is widely applied in predicting long-term trends in meteorology, hydrology, and ecological systems (Li et al., 2025; Bashir et al., 2020; Wang J. Y. et al., 2025). According to the range of H values, the persistence characteristics of the time series evolution trend could be determined: when 0.5 < H < 1, the future change trend of the time series is in the same direction as the historical phase, showing persistent characteristics; when H = 0.5, there is no long-term correlation between sequence values, and the future trend contains uncertainty; when 0 < H < 0.5, the sequence shows anti-persistent characteristics, and the future evolution direction is opposite to the historical trend. On this basis, combined with the Sen trend analysis, the future change trends of eRSEI could be divided into five categories: continued increase (β > 0.0005, H > 0.5), reversed to increase (β < -0.0005, H < 0.5), relatively stable (|β|<0.0005), reversed to decrease (β > 0.0005, H < 0.5), and continued decrease (β < -0.0005, H > 0.5).

The GDM, utilizing optimal parameters suggested by Song et al. (2020), was employed to systematically examine the driving mechanism of IMWSA eRSEI through three modules: factor detection, interaction detection, and risk detection. (1) Factor detection: Factor detection is the core component of GDM, used to determine the degree to which driving factors explain the spatial differentiation of the dependent variable, represented by the q statistic, with a numerical range of [0,1]. A higher q value indicates a greater explanatory power of that driving factor for eRSEI. (2) Interaction detection: By calculating the interaction parameter q between two factors, the explanatory power of their combined influence on the spatial distribution of eRSEI can be assessed, revealing both the type and strength of their interaction (Table 4). (3) Risk detection: This module identifies whether significant differences exist in the mean values of eRSEI across different categories defined by driving factors. It helps determine the specific value ranges or classes most associated with high EEQ levels, thereby revealing optimal ecological conditions (Wang R. B. et al., 2025).

Partial correlation analysis enables a systematic examination of the direction and strength of the relationships between the target variable and multiple driving factors. Building on the results of GDM, it further reveals the promoting or suppressing effects of meteorological changes on eRSEI. The formula for calculating the g-order partial correlation coefficient is given in Equation 6 (Wan et al., 2023): r i j , l 1 l 2 ⋯ l g = r i j , l 1 l 2 ⋯ l g − 1 − r i l , g l 1 l 2 ⋯ l g − 1 r j l , g l 1 l 2 ⋯ l g − 1 1 − r i l , g l 1 l 2 ⋯ l g − 1 2 1 − r j l , g l 1 l 2 ⋯ l g − 1 2 (6) where   r i j , l 1 l 2 ⋯ l g is the partial correlation coefficient between variables i and j under the condition of controlling variables l 1 l 2 ⋯ l g , with a numerical range of [-1,1]. A larger absolute value indicates a stronger correlation between the variables.

The temporal variation characteristics of each component indicator and the eRSEI in IMWSA are shown in Figure 3 From 2000 to 2024, NDVI exhibited an highly significant upward trend (P < 0.01), with an increase rate of approximately 0.020/10a. WET exhibited a non-significant upward trend at a rate of 0.010/10a (P > 0.05). Both NDBSI and SI-T displayed significant decreasing trends (P < 0.05), with rates of −0.020/10a and −0.070/10a, respectively. No significant trend was observed for LST. Under the combined influence of these indicators, eRSEI showed an overall increasing trend, with a rate of approximately 0.006/10a (P > 0.05). The temporal pattern could be divided into two phases: a decreasing trend from 2000 to 2010 at a rate of −0.005/10a (P > 0.05), followed by an increasing trend after 2010 at a rate of 0.009/10a (P > 0.05). Although the overall trend was not statistically significant, the three highest values in the time series all occurred after 2015, with a maximum of 0.236 recorded in 2018, suggesting an overall improvement in recent years.

To characterize the spatial distribution of eRSEI in IMWSA across different time periods, eRSEI images from 2000, 2005, 2010, 2015, 2020, and 2024 were selected and classified into five grades: poor, fair, moderate, good, and excellent (Figure 4). A clear spatial pattern emerged during the study period, as eRSEI values were generally higher in the east and declined toward the west. The poor-grade dominated most of the region, whereas good and excellent grades together accounted for less than 4% of the total area. Spatially, poor-grade areas were primarily distributed in Alxa, northern Bayannur, and the Kubuqi Desert in northern Ordos. Fair-grade areas were mainly found in central and eastern Bayannur, Wuhai, and around the Mu Us Desert in southern Ordos. The loess hilly region in eastern Ordos largely fell into the moderate grade. Good and excellent grades were concentrated in the Hetao Plain in southern Bayannur and in farmland and dense forest-grassland areas along the Yellow River in northern Ordos. Temporally, from 2000 to 2010, poor-grade areas expanded significantly, particularly in the Badain Jaran Desert and western Bayannur. Between 2010 and 2015, these areas slightly contracted. After 2020, regional divergence became more apparent: poor-grade areas increased in the Badain Jaran Desert but continued to decline in western Bayannur. Between 2020 and 2024, excellent-grade areas expanded markedly in the Hetao Plain. Meanwhile, the EEQ grade structure of the Mu Us Desert gradually shifted from a dominance of fair grades to a more balanced distribution of fair and moderate grades.

As shown in Figure 5, the statistical analysis of eRSEI grades from 2000 to 2024 indicates that the study area has been predominantly characterized by poor and fair grades, which together accounted for 81%–95% of the total area. Among these, the proportion of poor-grade areas remained relatively stable at approximately 63%. Fair-grade areas showed a marked decline, decreasing from 33% in 2000 to 23% in 2024, with a highly significant downward trend (P < 0.01). The proportion of moderate-grade areas fluctuated between 3% and 14%, displaying more variability than other grades, though they remained around 10% over the past 3 years. Good-grade areas exhibited the most pronounced expansion (P < 0.01), with a 2.5-fold increase in 2024 compared to 2000. Excellent-grade areas consistently represented the smallest proportion among all grades, remaining below 0.1% in more than half of the years. However, after 2018, a noticeable increase was observed, reaching 1.1% by 2024. Overall, the EEQ structure of the region exhibited a transitional pattern, characterized by a reduction in low-grade areas and expansion of moderate to high-grade areas, indicating a general improvement in EEQ of IMWSA.

Pixel-level spatial changes of eRSEI in the IMWSA from 2000 to 2024 are shown in Figure 6a1a2. The results indicated that eRSEI values increased in approximately 33.67% of the total area, with 9.52% showing statistically significant increase (P < 0.05). Decreasing eRSEI values were observed in 39.96% of the region, though only 1.69% exhibited a significant decrease (P < 0.05), while the remaining areas displayed no obvious change. Spatially, increases in eRSEI were primarily observed in most areas of Ordos, northern Bayannur, and the Ejina River Basin in central Alxa. Decreases were mainly concentrated in western and southern Alxa, as well as parts of northern Bayannur, with smaller patches also occurring in eastern Ordos. The major desert and sandy systems within the region exhibited differentiated evolutionary patterns. The most pronounced increase in eRSEI values was observed in the Mu Us Sandy Land, while the southern Tengger Desert and eastern Kubuqi Desert also exhibited upward trends. Meanwhile, eRSEI values generally decreased in the continuous sandy areas including the Badain Jaran Desert, Ulan Buh Desert, northern Tengger Desert, and western Kubuqi Desert, suggesting these regions have continued to undergo EEQ deterioration.

The coefficient of variation of eRSEI in the study area ranged from 0.02 to 0.97, with a mean value of 0.17 (Figure 6b1b2). Low variation was observed in 7.46% of the area, primarily concentrated in the Hetao Plain. Relatively low variation was identified in 68.21% of the area, exhibiting a continuous spatial distribution across the entire region. Moderate variation was present in 21.83% of the area, mainly distributed across most of Alxa, extending into northern Bayannur and western Ordos. Relatively strong and strong variations were found in only 2.50% of the area, primarily concentrated in the Badain Jaran Desert and its surrounding regions.

The average H value of the study area was 0.38 (Figure 6c1c2), indicating an overall anti-persistence characteristic. In conjunction with Figure 6a1, areas projected to experience an increase in eRSEI accounted for 32.60% of the total area, with 2.20% classified as areas of continued increase, scattered across the Hetao Plain and Ejina River Basin, and 30.40% were identified as reversed to increase areas, mainly concentrated in the central part of the study area, suggesting ecological restoration potential in these regions. Areas projected to experience a future decrease in eRSEI accounted for 30.21%, with 2.1% classified as areas of continued decrease, sporadically distributed across the Tengger Desert, Ulan Buh Desert, and surrounding regions, and 28.11% were identified as reversed to decrease areas, mainly covering most of Ordos, the northern Hetao Plain in Bayannur, and the Ejina Basin in Alxa, which are currently high-value zones requiring targeted prevention and management efforts to mitigate ecological deterioration risks.

The spatial differentiation of eRSEI in the IMWSA from 2000 to 2024 was analyzed using GDM. As shown in Figure 7, the factor detection results indicated that, at a 1% significance level (P < 0.01), all selected driving factors significantly explained the spatial heterogeneity of EEQ in the study area. The factors’ explanatory power (q-values), listed in descending order, were as follows: land use type (0.734) > precipitation (0.534) > temperature (0.405) > soil type (0.295) > solar radiation (0.051) > elevation (0.041). Among these, land use type exhibited the highest q-value and was identified as the primary driving factor influencing eRSEI in the IMWSA. Precipitation and temperature also showed significant effects. Although solar radiation and elevation had relatively low q-values, their explanatory power was substantially enhanced through interaction with other factors.

The interaction detection results indicated that the combined influence of any two factors exceeded that of each individual factor, exhibiting both bivariate enhancement and nonlinear enhancement effects. The highest interaction explanatory power was observed between land use type and precipitation (q = 0.827), suggesting that the synergistic effect of human activities, represented by land use, and climatic factors could significantly improve the explanatory power of eRSEI. In addition, the interaction q-values between land use type and soil type, temperature, solar radiation, and elevation, as well as between solar radiation and precipitation, all exceeded 0.730. This highlighted the important role of multi-factor coupling in shaping the spatial distribution of eRSEI.

Based on the risk detection module, the optimal ranges or categories of each driving factor in relation to eRSEI were analyzed. The results indicated a positive correlation between eRSEI and precipitation within the study area, with eRSEI increasing as precipitation rises. The highest eRSEI values were observed when summer precipitation ranged from 249 to 307 mm. Elevated eRSEI levels were sustained when summer temperatures averaged 16.4 °C–22.5 °C, accompanied by solar radiation ranging from 1,570 to 1640 MJ m^-2. The highest eRSEI values occurred at elevations between 1,600 and 3,300 m. Among different soil types, semi-hydromorphic soils corresponded to the highest eRSEI values. Regarding land use types, woodland exhibited significantly higher eRSEI values than unused land and grassland, highlighting the positive impact of vegetation cover on ecological quality.

Partial correlation analysis (Figure 8) revealed that eRSEI in IMWSA was generally positively correlated with temperature, precipitation, and solar radiation, with mean partial correlation coefficients of 0.02, 0.26, and 0.03, respectively. Among these, precipitation exhibited the strongest correlation with eRSEI. Areas showing a positive correlation between eRSEI and temperature accounted for 53.78% of the region, primarily distributed in the northwestern and southeastern parts of the study area. Areas with significant and highly significant positive correlations accounted for 2.62%, mainly concentrated in the Hetao Plain, with scattered patches in the central Mu Us Sandy Land and the northern edge of Alxa. In contrast, negative correlation with temperature was found in 46.22% of the region, among which 1.37% showed significant or highly significant negative correlation, primarily distributed in the Badain Jaran Desert and Tengger Desert. Partial correlation between eRSEI and precipitation showed a general east-west gradient. Over 88% of the study area showed positive correlation with precipitation, including 14.96% with significant and 9.41% with highly significant positive correlation, mainly distributed in southern Bayannur, the Mu Us Sandy Land, and the Tengger Desert. Regions showing a negative correlation with precipitation were primarily located around the Badain Jaran Desert; however, less than 0.1% exhibited statistical significance. Regarding solar radiation, positive and negative correlation with eRSEI were observed in 55.34% and 44.66% of the study area, respectively. Areas with significant and highly significant positive correlations were scattered in the central part of the study area, accounting for approximately 2.64%. Conversely, areas with significant and highly significant negative correlations were mainly concentrated around the Ejina River Basin, accounting for about 1.24%.

According to the absolute values of the partial correlation coefficients (Figure 9), precipitation is identified as the most critical factor influencing eRSEI variation in IMWSA, dominating changes in 58.07% of the area. Temperature ranked second, with a dominant influence on 22.42% of the region, while solar radiation contributed the least, primarily affecting 19.51% of the area. From a spatial perspective, precipitation exerted a predominantly positive influence on eRSEI variation across most of the study area. However, in regions with severe desertification and extremely low vegetation cover, short–term heavy rainfall can induce soil erosion, leading to potential negative ecological impacts. Areas dominated by temperature and solar radiation were mainly concentrated in central Alxa, whereas in agricultural zones such as the Hetao Plain, where soil moisture conditions are regulated by human activities, EEQ was more sensitive to changes in temperature.

Land use change serves as a direct reflection of both the direction and intensity of human activities in modifying the natural environment. As shown in Table 5, land use changes in IMWSA from 2000 to 2023 accounted for 9.54% of the total area. The proportions of grassland converted into other land types and converted from other land types were nearly balanced, both around 4.55%. Cropland experienced a 0.96% area loss but gained 1.93% through conversion, leading to a net increase of about 0.97%. Unused land decreased, as 3.97% was converted to other uses while only 2.63% was gained through conversion. Although the areas of woodland and built-up land showed an overall increasing trend, their spatial proportions remained relatively low within the study area.

In terms of the impact of land use type conversion on EEQ, most land transitions had a positive effect on eRSEI, except for the conversion of cropland and grassland into unused land, which led to a decline in eRSEI. Specifically, the conversion from unused land to cropland resulted in the most significant improvement in eRSEI, with an increase of 0.28. The conversions from grassland to cropland (0.19) and from built-up land to cropland (0.14) also notably enhanced eRSEI. During the study period, the average eRSEI in areas undergoing land use change increased by 0.09, while that in stable land use areas rose by only 0.02. This difference confirms that land use transformation played a promotive role in enhancing regional EEQ.