Introduction

Vegetation is a key hub connecting the material cycle, energy flow and information transmission of terrestrial ecosystems1,2, and its coverage changes directly reflect the dynamics of terrestrial ecosystems3,4. In arid and semi-arid regions that are highly sensitive to climate and human activities, vegetation indices are particularly important for studying hydrological processes, ecosystem functions and regional environmental changes5,6. In these ecosystems, vegetation dynamics not only serve as indicators of climate change but also form a close two-way feedback with soil physical and chemical properties and water availability. Therefore, quantitatively characterizing the spatio-temporal variation pattern of vegetation and revealing its climate-soil-water driving mechanism have significant scientific and practical significance for ecological protection.

In remote sensing monitoring, NDVI has a good correlation with biomass and leaf area index, and can well reflect the surface vegetation coverage. This can be used to effectively characterize vegetation activities and productivity, and is applicable to representing changes in surface vegetation coverage7,8. However, the sensitivity of NDVI to the soil background and its saturation trend in areas with high vegetation coverage lead to a decrease in its sensitivity to changes in vegetation density9,10,11. Therefore, Camps Valls et al.12 utilized the principles of machine learning and applied the kernel method theory to the extraction and calculation of NDVI, proposing a new vegetation index, kNDVI. The kNDVI index aims to address the saturation issue of traditional NDVI in areas with high vegetation coverage, to more accurately reflect the biophysical and physiological states of vegetation13.Multiple studies have shown that, compared with NDVI, kNDVI performs better in capturing seasonal dynamics, gross primary productivity (GPP), and solar-induced chlorophyll fluorescence (SIF), with particularly pronounced advantages in high-coverage and arid regions14,15. Furthermore, its sensitivity to biophysical processes is higher in tropical forests and restoration ecosystems12. At present, kNDVI has been widely applied in monitoring vegetation cover, assessing growth status, and analyzing vegetation responses to climate change and land-use change. For example, Liu et al.16 used kNDVI to reveal the relationship between vegetation restoration and land use in the Yellow River Basin; Feng et al.17 combined Mann-Kendall mutation (M-K) test, Theil-Sen slope analysis and Mann–Kendall test analysis to study the spatio-temporal variation characteristics of global grassland kNDVI. However, existing studies rarely carry out systematic benchmark comparisons between NDVI and kNDVI in arid and semi-arid regions, particularly in relation to dynamic range, soil-background effects, high-coverage saturation and coupling with water processes, leading to potentially biased trend detection and attribution results.

At present, studies on vegetation change mostly focus on overall trend analyses of kNDVI. Among them, Theil–Sen slope estimation is a commonly used non-parametric method for quantifying the long-term trend of kNDVI time series, offering strong robustness to outliers and skewed distributions18. However, this method can only reflect the magnitude of trend change and cannot determine its statistical significance, which limits its interpretability. To address this limitation, studies often combine the Mann–Kendall (MK) test with the Theil–Sen slope estimation method to quantify both the trend magnitude and its statistical significance. Compared with statistical methods relying on parametric assumptions (such as ordinary least squares), this combined approach is more robust when handling time series with outliers or non-normal distributions, and is therefore widely applied in vegetation trend analysis19,20. Recently, a robust trend analysis framework integrating Theil–Sen estimation, contextualized MK tests, and false discovery rate (FDR) control has been developed to improve the accuracy and reliability of large-scale remote sensing trend estimation21.However, focusing solely on the interannual variation of kNDVI may overlook short-term disturbances embedded in the time series. To capture vegetation dynamics more precisely, it is necessary to monitor and evaluate abrupt changes within the time series22. Various methods have been developed for vegetation change detection, including disturbance and recovery trend detection based on Landsat imagery23 and the Vegetation Change Tracker24. Time-series segmentation algorithms such as LandTrendr can identify sudden disturbances, sustained degradation, and recovery phases within long time series, and have been widely applied in ecological project evaluation and risk monitoring23. Due to the diversity and uncertainty of vegetation change, only high-temporal-resolution remote sensing data can fully describe the complete vegetation change process over short time scales25. Dense satellite time series thus form the basis for dynamic vegetation cover monitoring6, and time-series change analysis has attracted increasing attention. This approach has been increasingly applied to kNDVI time series for trend detection and breakpoint identification, time-series smoothing and interpolation, and land surface phenology analysis26. Therefore, relying solely on monotonic trends over the entire period tends to obscure structural shifts and stage-specific divergent changes, and makes it difficult to assess whether a trend is persistent. This underscores the urgent need for an integrated diagnostic framework that links “abrupt change detection–segmented evaluation–persistence interpretation.”

Inner Mongolia, located in northern China along the border, covers a vast territory with varied topography and diverse landscapes, and is among the most sensitive regions in the world to environmental changes14. The region has an arid and semi-arid climate, and over the past 40 years, its rate of temperature increase has far exceeded the global average27,28. Studies indicate that its warming magnitude is about two to three times the global average29, accompanied by decreasing precipitation and increasing groundwater depth, which have heightened the risk of vegetation degradation. Moreover, since 2000, the region has entered a stage of rapid socio-economic development. Groundwater depth, climate, soil conditions, and human activities are rapidly reshaping the structure and function of vegetation systems at multiple levels, with significant implications for ecology and sustainability. Against this backdrop, water availability is not only a key limiting factor for vegetation growth but also an important driver of soil structure and nutrient cycling, while soil texture, topography, and land-use practices regulate the accessibility of water to vegetation, forming a complex coupled network30,31. However, few regional-scale studies quantitatively compare groundwater, precipitation, temperature, soil physicochemical properties, topography and land use, or rank their relative dominance. Most research still focuses on single-factor correlations, which limits the ability to distinguish primary from secondary drivers within spatial heterogeneity.

In summary, this study first conducts a systematic comparison between NDVI and kNDVI to select the most suitable primary index of vegetation dynamics. Based on this, an integrated “trend–abrupt change–persistence” diagnostic framework is established: sequential MK tests are used to identify abrupt change points and segment the series; within each segment, Theil–Sen estimation combined with the MK test is applied to assess trend magnitude and significance; persistence is then interpreted using the Hurst exponent. Finally, pixel-wise Pearson correlation and Geodetector analyses are employed to quantitatively compare and spatially rank the explanatory power of factors including groundwater, precipitation, temperature, soil physicochemical properties, topography, and land use, thereby extracting the dominant factor patterns underlying vegetation spatial heterogeneity in the region. This multi-factor framework aims to reveal vegetation–soil–water relationships and their spatiotemporal variations in arid and semi-arid regions, providing quantitative evidence to support ecological restoration and water resource management.

Materials and methods

Study area

The Inner Mongolia Autonomous Region is located in northern China along the border (97° 12′–126° 04′ E, 37° 24′–53° 23′ N), covering a total area of 1.183 × 106 km2, which accounts for 12.3% of China’s total land area, making it the third-largest provincial-level administrative unit in the country. The region is governed by a temperate continental monsoon climate, with pronounced regional and seasonal variations: the annual precipitation is around 375 mm, decreasing from northeast to southwest, forming a distinct moisture gradient; mean annual temperature ranges from − 4.65 to 9.14 °C, shows an opposite trend, increasing from the northeast to the southwest; and groundwater depth is between 8 and 165 m generally increases along this gradient of hydrothermal conditions. Soil types display clear spatial differentiation, transitioning from black soils, dark brown loams, and black calcareous soils in the east, to chestnut calcareous soils, brown loams, gray calcareous soils, sandy soils, and finally gray desert soils toward the west. Vegetation types follow the precipitation gradient from forests in the east, to grasslands in the central zone, and deserts in the west. Grasslands can be further classified by climatic zone into meadow steppe, typical steppe, and desert steppe32 (Fig. 1). This coordinated gradient of water–energy, soil, and vegetation provide a natural spatial laboratory for analyzing vegetation–soil–water coupling mechanisms at the regional scale.

Fig. 1
figure 1

The map of the geographic location of study area (GS(2022)4301). This base map is publicly available and has been included in the title of Fig. 1. This map was created by me using the open-source software QGIS 3.40.0.

Full size image

Data sources

Remote sensing data

Landsat imagery (L5 TM, L7 ETM+, L8 OLI) from January 2000 to December 2024 was acquired via the Google Earth Engine (GEE) platform, with QA_PIXEL/CFMask applied for cloud masking and scenes with cloud cover greater than 20% excluded. To enhance comparability, all data were standardized to a 30 m projection and resolution (bilinear resampling for continuous rasters and nearest-neighbor resampling for categorical rasters), and annual (or growing season) median composites were generated. The data sources and key parameters and some influencing factors are shown in Table 1; Fig. 2.

Fig. 2
figure 2

kNDVI drivers and spatial distribution in Inner Mongolia. This base map is publicly available and has been included in the title of Fig. 1. This map was created by me using the open-source software QGIS 3.40.0.

Full size image
Table 1 Data used in this study.
Full size table

Soil datasets

To cover different land-use types, sampling sites were laid out in the summer of 2024 at an approximate density of one site per 100,000 km2 (Fig. 3). Surface soil samples (0–30 cm) were collected using a stainless-steel auger; three subsamples were taken at each site and combined into a single composite sample, yielding a total of 363 samples. Laboratory measurements and methods were as follows: pH was determined in a 2.5:1 water-to-soil suspension33; total nitrogen (TN) was measured using the semi-micro Kjeldahl method34; total phosphorus (TP) was determined via the NaOH fusion–molybdenum-antimony colorimetric method35; total potassium (TK) was determined via the NaOH fusion–flame photometry method36; soil organic matter (SOM) was determined using the potassium dichromate oxidation–external heating method37; available phosphorus (AP) was extracted with NaHCO₃ and determined using a UV–visible spectrophotometer38; available potassium (AK) was extracted with NH₄OAc and determined using flame photometry39; and alkali-hydrolyzable nitrogen (AN) was determined using the alkali-hydrolysis diffusion method40. All analyses were conducted in accordance with the corresponding standard methods and under a unified quality control protocol to ensure data comparability and reliability.

Fig. 3
figure 3

Layout of the sampling points in the study area. This base map is publicly available and has been included in the title of Fig. 1. This map was created by me using the open-source software QGIS 3.40.0.

Full size image

To meet the normality and variance stability assumptions of Kriging, all variables were log-transformed, and semi variograms were constructed and fitted with the optimal model (Spherical, Exponential, or Gaussian) for each variable. Ordinary Kriging interpolation was then performed accordingly41. Cross-validation results showed that the mean error (ME) was close to 0, the root mean square standardized error (RMSSE) was approximately 1, and the RMSE was similar to the ASE, indicating that the interpolation exhibited no significant bias and the uncertainty estimates were reasonable (Table 2).

Table 2 Semi-variance model parameters of ordinary kriging method.
Full size table

Application of kNDVI

The definition of NDVI is given in Eq. (1), and the kernel form of kNDVI is presented in Eqs. (2)–(4). Under a radial basis kernel with a specified σ, Eqs. (2)–(4) can be approximated as a nonlinear mapping of NDVI; in this study, the empirical formulation in Eq. (5) was adopted to facilitate large-scale computation. Pixel-wise NDVI values obtained from GEE were transformed to construct a kNDVI time series dataset for Inner Mongolia covering 2000–2024. Comparisons between kNDVI and NDVI were conducted annually, including analyses of spatial distribution, dynamic range, statistical characteristics, and saturation behavior in high-coverage areas, as well as correlation and explanatory power assessments with water-related variables (precipitation, soil moisture, and groundwater depth), in order to validate the rationale for using kNDVI.

$$\:NDVI=\frac{n-r}{n+r}$$
(1)
$$\:kNDVI=\frac{k(n,n)-k(n,r)}{k(n,n)+k(n,r)}$$
(2)

Here, n denotes the near-infrared (NIR) band, r denotes the red band, and k represents the kernel function, which is used to measure the similarity between the NIR and red bands. The kernel function is typically implemented using a radial basis function (RBF), expressed as:

$$\:k\left(a,b\right)=\text{exp}\left(\frac{-({a-b)}^{2}}{{2\sigma\:}^{2}}\right)$$
(3)

Here, σ is the length-scale parameter, proportional to the mean reflectance of the near-infrared and red bands in the acquired remote sensing imagery. Then, using Eq. (3), k(n, n) and k(n, r) are calculated and substituted into Eq. (2). Through mathematical transformation, the expression is converted into the form of a hyperbolic tangent function:

$$\:kNDVI=\text{tanh}\left({\left(\frac{n-r}{2\sigma\:}\right)}^{2}\right)$$
(4)

If σ = 0.5 (n + r), a good compromise between accuracy and simplicity can be achieved, allowing for more efficient pixel-wise computation. The value of σ has a significant impact on the smoothness of the kernel function: a larger σ produces a smoother response, effectively reducing noise but potentially causing loss of detail; conversely, a smaller σ preserves more detail but is more sensitive to noise. By setting σ to half of the average of n and r, an appropriate balance between smoothness and detail preservation is attained. This simple linear combination not only simplifies computation but also improves processing efficiency, particularly for large-scale remote sensing datasets. Moreover, experiments have shown that this specific σ value performs well across a variety of scenarios, and it is therefore adopted as the default setting42. Accordingly, Eq. (4) can be simplified as:

$$\:kNDVI=\text{tanh}\left({NDVI}^{2}\right)$$
(5)

Time-series trend assessment and abrupt change detection methods

This study adopts an integrated workflow of “trend magnitude and significance + abrupt change detection and phase segmentation.” First, the Theil–Sen method is used to estimate the annual rate of change (magnitude/direction), and the Mann–Kendall (MK) test is applied to assess statistical significance. Sequential MK (UF/UB) analysis is then employed to identify abrupt change years and segment the time series, thereby avoiding the masking of phase-specific divergent changes by full-period averages.

Trend magnitude and significance: Theil–Sen + Mann–Kendall

The Theil–Sen slope estimation is a robust, nonparametric method that does not rely on distributional assumptions, making it suitable for handling missing values in time series and reducing the influence of data distribution on results. Its strong resistance to outliers ensures reliable performance across diverse datasets, thereby enhancing the scientific credibility of the findings18. In this study, the Theil–Sen median slope estimator is applied to quantify the temporal trends of kNDVI in Inner Mongolia from 2000 to 2024. The slope is calculated using Eq. (6):

$$\:\beta\:=median\frac{{kNDVI}_{j}-{kNDVI}_{i}}{j-i}(2000\le\:i<j\le\:2024)$$
(6)

where \(\:{kNDVI}_{j}\:\text{a}\text{n}\text{d}\:{kNDVI}_{i}\) represent the kNDVI values in years j and i, respectively, and Median denotes the median function. A β value greater than 0 indicates an improving kNDVI trend, whereas a negative β value indicates degradation.

$$Z = \left\{ {\begin{array}{*{20}l} {(S - 1)/\sqrt {Var\left( S \right)} ,} \hfill & {S > 0} \hfill \\ 0 \hfill & {S = 0} \hfill \\ {(S + 1)/\sqrt {Var\left( S \right)} ,} \hfill & {S < 0} \hfill \\ \end{array} } \right.$$
(7)
$$\:S=\sum\:_{i=1}^{n-1}\sum\:_{j=i+1}^{n}sgn({kNDVI}_{j}-{kNDVI}_{i})$$
(8)
$$\:sgn({kNDVI}_{j}-{kNDVI}_{i})=\left\{\begin{array}{c}1({kNDVI}_{j}-{kNDVI}_{i}>0)\\\:0({kNDVI}_{j}-{kNDVI}_{i}=0)\\\:-1({kNDVI}_{j}-{kNDVI}_{i}<0)\end{array}\right.$$
(9)
$$\:Var\left(S\right)=n(n-1)(2n+5)/18$$
(10)

where n is the length of the time series, sgn denotes the sign function, S is the test statistic for the Mann–Kendall test, and Var(S) is the variance of S. Z represents the standard normal variable, with significance assessed at α = 0.05. When |Z| ≥ 1.96, the change in kNDVI is considered statistically significant. This calculation is performed using MATLAB.

Sequential MK (UF/UB) test for abrupt change detection and phase segmentation

The MK abrupt change test is a nonparametric method that does not require a specific data distribution and can objectively reveal overall change trends without being affected by outliers. The basic principle is to identify abrupt change years by locating the intersection points between the forward sequence statistic (UF) and the backward sequence statistic (UB).

For a time series X with n observations, the forward sequence is constructed as follows:

$$S_{K} = \sum\limits_{{i = 1}}^{k} {r_{i} r_{i} } = \left\{ {\begin{array}{*{20}l} 1 \hfill & {x_{i} > x_{j} } \hfill \\ 0 \hfill & {else} \hfill \\ \end{array} } \right.\quad j = {\text{1,2}}, \ldots ,i$$
(11)

Here, ri represents the cumulative count of xⱼ values (for 1 < j < i) that are smaller than xi. Under the assumption that the time series is random and independent, the\(\:{UF}_{k}\) statistic is defined as:

$$\:{UF}_{k}=\frac{{S}_{K}-E\left({S}_{K}\right)}{\sqrt{Var\left({S}_{K}\right)}}\:k=\text{1,2},.\:.\:.\:,n$$
(12)

where \(\:{UF}_{1}\)=0,\(\:E\left({S}_{K}\right)\) and \(\:Var\left({S}_{K}\right)\) denote the mean and variance of the cumulative sum Sₖ under the assumption of independent and identically distributed variables. The variables are x1, x2, …, xn, and are calculated as follows\(\:(E\left({S}_{K}\right)=\frac{n\left(n-1\right)}{4})\):

$$\:Var\left({S}_{K}\right)=\frac{n(n-1)(2n+5)}{72}$$
(13)

\(\:{UF}_{i}\) is the normalized statistic derived from the time series \(\:{x}_{1},{x}_{2},\dots\:{x}_{n}\). Using xn, the statistical sequence is calculated, and at a significance level α, the standard normal distribution table is referenced. If \(\:\left|{UF}_{i}\right|\)>\(\:{U}_{\alpha\:}\), the series exhibits a significant trend change. For the reverse time series xn, x(n−1), …, x1, the same procedure is repeated, where \(\:{UB}_{k}\)=\(\:{-UF}_{k}\) for k = n, n − 1,…1,\(\:{UB}_{1}\) = 0.

Hurst index

The Hurst method43 quantitatively evaluates the degree of dependence and continuity in long-term time series data, reflecting the extent to which past information influences future trends. The specific formula is given as follows:

$$\:\frac{R}{S}={\left(cm\right)}^{H}$$
(14)

where R is the range (extreme deviation), S is the standard deviation, c is a constant, and the observations are divided into n subsequences \(\:{kNDVI}_{i}\), where i = 1,2,.,n. m is any positive integer with 0 < m < n, H is the Hurst exponent, and the formula for calculating the range R is as follows:

$$\:R=maxX(t,m)-minX(t,m)$$
(15)

where X(t) is the cumulative deviation, calculated as follows:

$$\:X\left(t\right)=\sum\:_{i=1}^{m}\left({kNDVI}_{i}-\frac{1}{m}\sum\:_{i=1}^{m}{kNDVI}_{i}\right)$$
(16)

where, 1 < t < m; standard deviation S is calculated as follows:

$$\:S=\sqrt{\frac{1}{m}{\sum\:_{i=1}^{m}({kNDVI}_{i}-\frac{1}{m}\sum\:_{i=1}^{m}{kNDVI}_{i})}^{2}}$$
(17)

The Hurst index (H) ranges from 0 to 1 and indicates the degree of persistence or anti-persistence in kNDVI trends. When 0 < H < 0.5, the series is anti-persistent, meaning future trends are likely to be opposite to past trends; the closer H is to 0, the stronger the anti-persistence. When H = 0.5, the series behaves randomly with no discernible trend. When 0.5 < H < 1, the series is persistent, meaning future trends tend to follow past trends; the closer H is to 1, the stronger the persistence44. For analysis, H values were classified into four categories: strong anti-persistence (< 0.35), weak anti-persistence (0.35–0.5), weak persistence (0.5–0.65), and strong persistence (> 0.65). The future trajectory of kNDVI can thus be characterized by combining the Hurst index with the observed trend direction (Table 3).

Table 3 Classification of future vegetation change trends in the study area.
Full size table

Analysis of driving factors of vegetation change

Pearson correlation analysis and significance test

The Pearson correlation coefficient is a statistical measure that describes the strength and direction of the linear relationship between two variables. Using MATLAB R2023a, pixel-wise correlations were calculated between kNDVI and groundwater depth(Monthly in 2024), precipitation, and temperature (2000–2024). The calculation formula is as follows:

$$\:r=\frac{\sum\:_{i=1}^{n}({x}_{i}-\overline{x}\left)\right({y}_{i}-\overline{y})}{\sqrt{{\sum\:_{i=1}^{n}({x}_{i}-\overline{x})}^{2}}\sqrt{{\sum\:_{i=1}^{n}({y}_{i}-\overline{y})}^{2}}}$$
(18)

In the formula, xi is the value of the influencing factor in year i; \(\overline{x}\) is the 25-year mean value of that influencing factor; yi is the kNDVI value in year i; ȳ is the 25-year mean value of kNDVI; n is the number of years; and r is the Pearson correlation coefficient between kNDVI and the influencing factor. The value of r ranges from − 1 to 1, with its absolute value closer to 1 indicating a stronger correlation. Significance testing is performed using an F-test. When the p-value satisfies 0.05 ≤ P ≤ 1, 0.01 ≤ P < 0.05, and P < 0.01, the correlation between kNDVI and the influencing factor is considered non-significant, significant, and highly significant, respectively.

Geodetector

Considering the study area’s conditions and data availability, key factors closely related to kNDVI—land use and land cover (LULC), available nitrogen (AN), available potassium (AK), available phosphorus (AP), total nitrogen (TN), total potassium (TK), total phosphorus (TP), soil organic matter (SOM), pH, elevation (DEM), slope (SLOP), and relief (REL)—were selected to investigate the driving forces behind the spatiotemporal evolution of kNDVI.

Geodetector analysis was applied to identify the factors underlying the observed patterns, quantifying the explanatory power of each independent variable for the dependent variable. To determine the dominant factors influencing the spatial variation of kNDVI, a single-factor detector was used to assess the spatial heterogeneity associated with each factor. The formula is:

$$\:\text{q}=1-\frac{\sum\:_{\text{h}=1}^{\text{L}}{\text{N}}_{\text{h}}{{{\upsigma\:}}^{2}}_{\text{h}}}{\text{N}{{\upsigma\:}}^{2}}\:$$
(19)

where q represents the explanatory power for kNDVI, h = 1, 2, …, L denotes the stratification of the variable, Nₕ and N are the sample sizes of class h and the entire region, respectively, and σ2 and\(\:{{{\upsigma\:}}^{2}}_{\text{h}}\) are the variances of kNDVI for the whole region and for class h, respectively. The value of q ranges from 0 to 1, reflecting the degree of spatial differentiation of the dependent variable. A larger q value indicates stronger explanatory power for the kNDVI variation in the study area45.

Results

NDVI–kNDVI index suitability and external validation

Spatiotemporal comparison

To assess the suitability of the two indices for the study area, vegetation spatial distributions for the years 2000, 2005, 2010, 2015, 2020, and 2024 were compared year by year (Table 4; Fig. 4). Overall, vegetation growth in the region has remained at a relatively low level over the past 25 years: according to kNDVI statistics, 80.49% of the area had pixel values < 0.5 (Fig. 4), indicating a widespread low-growth state. In contrast, NDVI statistics showed only 17.96% of the area in the 0–0.1 range, making it difficult to accurately reveal the true extent and severity of low-growth or stressed regions. Compared with NDVI, kNDVI exhibited a more dispersed value distribution and more pronounced spatial heterogeneity (Table 4), enabling clearer differentiation of low-growth zones and patches of varying levels, and thus more sensitively identifying and characterizing areas of poor vegetation growth. Overall, the comparison shows that kNDVI outperforms NDVI in dynamic range, resolution of low-growth areas, and expression of spatial differences, making it more suitable for vegetation condition monitoring and assessment in this region.

Fig. 4
figure 4

Spatial distribution of changes in NDVI and kNDVI from 2000 to 2024. This base map is publicly available and has been included in the title of Fig. 1. This map was created by me using the open-source software QGIS 3.40.0.

Full size image
Table 4 Comparison of variation coefficients of NDVI and kNDVI in the different years.
Full size table

To further validate the suitability of kNDVI compared with NDVI, several verification sites were selected in the study area in June 2024 (Fig. 5). High-resolution RGB orthophotos (spatial resolution < 5 cm) were acquired as reference data and compared with NDVI and kNDVI raster data for the corresponding areas. Each site included areas of low vegetation cover and exposed bare soil. The results showed that kNDVI values in low-coverage areas were closer to 0, allowing for more accurate identification of poorly growing or stressed vegetation patches, whereas NDVI values in the same areas tended to be higher, showing insufficient discrimination from bare soil backgrounds. Further pixel-based statistics within the sample sites revealed that the spatial coefficient of variation (CV) of kNDVI was consistently higher than that of NDVI, indicating greater sensitivity to fine-scale spatial differences and better agreement with orthophoto interpretation. Based on these comparisons, kNDVI was selected as the primary vegetation status indicator for subsequent analyses.

Fig. 5
figure 5

Remote sensing image data acquired in the study area. This base map is publicly available and has been included in the title of Fig. 1. This map was created by me using the open-source software QGIS 3.40.0.

Full size image

Seasonal baseline (Phenological Background)

Figure 6 shows the seasonal variation of the annual mean kNDVI in Inner Mongolia from 2000 to 2024. In winter, under cold and dry conditions, vegetation is generally dormant, and kNDVI drops to its lowest values. In spring, as temperatures rise and precipitation increases, kNDVI steadily increases, indicating the transition of vegetation from dormancy to active growth. During summer, photosynthetic activity is at its peak, and kNDVI reaches its annual maximum. In autumn, with declining temperatures and vegetation senescence, kNDVI decreases. Overall, the study area exhibits a clear phenological rhythm, with summer as the primary growing season. Linear trend analysis of seasonal extreme values shows that the summer peak kNDVI increased at a rate of 0.0012 yr−1 (R2 = 0.773), while the winter minimum increased at a rate of 0.0002 yr−1 (R2 = 0.772). Together, these results indicate that vegetation greenness in both the growing and non-growing seasons has shown an increasing trend, with the enhancement being more pronounced during the growing season.

Fig. 6
figure 6

Seasonal variations of the average kNDVI in Inner Mongolia from 2000 to 2024.

Full size image

Trend magnitude, significance and mutation segmentation

Sequential Mann–Kendall (MK) change-point detection and stage delineation

Based on the sequential MK change-point test, 2008 and 2016 were identified as significant breakpoints in Inner Mongolia’s kNDVI (Fig. 7a): the forward (UF) and backward (UB) statistics intersect in those years and lie within the ± 1.96 critical band, satisfying the α = 0.05 criterion. Accordingly, the record was segmented into 2000–2008, 2008–2016, and 2016–2024. The interannual series (Fig. 7b) shows that 2000–2008 was characterized by small-amplitude fluctuations with a gentle rise, as mean kNDVI increased from 0.216 to 0.261 (slope + 0.0056 yr−1). During 2008–2016, variability intensified with stagewise reversals: an overall increase in 2009–2012 (+ 0.0193 yr−1) was bracketed by declines in 2009–2010 (− 0.043 yr−1) and 2012–2016 (− 0.0085 yr−1), indicating a highly disturbed and unstable trend. From 2016 to 2024, the trajectory returned to growth, with mean kNDVI rising from 0.243 to 0.300 at + 0.0071 yr−1, exceeding the first stage. Overall, the structural breaks in 2008 and 2016 delineate an evolution of “slow increase → high volatility → stabilized increase”; the recent enhancement is more pronounced in the growing season, while the non-growing season also shows a modest uplift.

Fig. 7
figure 7

The variation of kNDVI in Inner Mongolia from 2000 to 2024: (a) MK mutation test, and (b) interannual variation.

Full size image

Using the Natural Breaks (Jenks) Classification Method, annual maximum kNDVI levels over the study area were classified into five vegetation-cover levels: low (0–0.09), low–moderate (0.09–0.22), moderate (0.22–0.36), moderate–high (0.36–0.52), and high (0.52–0.62) (Fig. 8). Overall, vegetation growth is relatively weak and exhibits a pronounced “higher in the east, lower in the west” pattern: the eastern region (Hulunbuir, Hinggan League, Tongliao, and Chifeng) is dominated by forests and typical grasslands with mostly moderate–high to high kNDVI, whereas the western region (Alxa League, Ordos, and Bayannur) is characterized by desert and desert–steppe with predominately low to low–moderate levels. This spatial differentiation aligns with the northeast-to-southwest precipitation decline across the region. By period, 2000–2008 had the largest share of low-cover area (38.81%); during 2008–2016, vegetation conditions improved relative to the first stage, with mean kNDVI reaching 0.3422; in 2016–2024, the overall pattern persisted and the mean further increased slightly to 0.3471, indicating a modest enhancement over the second stage.

Fig. 8
figure 8

Stage-wise spatial distribution and percentage of annual maximum kNDVI levels in Inner Mongolia (2000–2024). This base map is publicly available and has been included in the title of Fig. 1. This map was created by me using the open-source software QGIS 3.40.0.

Full size image

Full-period and segmented trend assessment with Theil–Sen + Mann–Kendall (TS + MK)

From 2000 to 2024, kNDVI in Inner Mongolia shows a gradual upward trend, with the most pronounced increases in the southeast and southwest (Fig. 9). After segmenting the record into three stages via the MK change-point test(Table 5), we find: during 2000–2008, areas with significant vegetation increase accounted for 11.69%; during 2008–2016, greening slowed, with the share of significant vegetation increase dropping to 8.27%, while 33.89% of the region experienced slight degradation; during 2016–2024, 11.96% of the area experienced significant vegetation increase with a ~ 4.8% patch of significant degradation emerged in the southwest. Despite the localized degradation in the third stage, the full-period balance is toward improvement, cumulative improvement covers ~ 47.52% of the area, indicating an overall greening of Inner Mongolia in the 21st century.

Fig. 9
figure 9

The interannual variation trend of vegetation growth in Inner Mongolia from 2000 to 2024. This base map is publicly available and has been included in the title of Fig. 1. This map was created by me using the open-source software QGIS 3.40.0.

Full size image
Table 5 The interannual variation trend of kNDVI of vegetation in inner Mongolia.
Full size table

Trend persistence and future direction: Theil–Sen × Hurst joint assessment

Based on the Hurst-exponent assessment of kNDVI persistence (Fig. 10a), most of Inner Mongolia exhibits anti-persistence (H < 0.5), accounting for 83.44% of the region. This indicates that the current vegetation trajectory is likely to reverse in a large number of areas rather than continue in the same direction. The low-H zones are mainly distributed along the border belt, implying high interannual variability and strong sensitivity to external climatic or human disturbances. In contrast, persistent areas (H > 0.5), which cover only 16.56%, are largely concentrated in southern Chifeng, southern Xilin Gol League and parts of Alxa League, suggesting a stronger “memory effect” and relatively stable, self-sustaining vegetation cover in these zones.

Fig. 10
figure 10

Spatial distribution of the Hurst index and future kNDVI trend levels in Inner Mongolia during 2000–2024. This base map is publicly available and has been included in the title of Fig. 1. This map was created by me using the open-source software QGIS 3.40.0.

Full size image

To further identify future vegetation change trajectories, we combined the trend sign (Theil–Sen estimator and MK test) with the Hurst-based persistence index to classify each pixel into five types (Fig. 10b). The results show that continuous improvement accounts for only 8.82% of the region, mainly distributed in Hulunbuir, Hinggan League and Tongliao, corresponding well with long-term restoration and soil-and-water conservation efforts. Continuous degradation areas (10.65%) are clustered in Alxa League, reflecting sustained water shortages and high disturbance pressure. The majority of the region (80.54%) falls into the basically unchanged category, indicating an overall pattern of near-term stability, although localised improvement–degradation mosaics persist. These spatial patterns reveal that improvement is more likely in the humid north-eastern forest–steppe belt, whereas continued degradation tends to occur in the western drylands where climatic and hydrological constraints remain strong.

Multi-factor drivers and spatial heterogeneity

Climate–water controls

Pixel-wise Pearson correlation analyses were carried out between kNDVI and the three environmental variables across Inner Mongolia (Fig. 11). The results show that kNDVI is predominantly positively correlated with groundwater depth, covering 64.62% of the region, but the proportion of significant pixels is relatively low (2.93% at p < 0.01 and 6.34% at p < 0.05), Positive correlations with precipitation are even more widespread (90.12%), including 28.68% highly significant and 20.54% significant areas, while negative relationships are only found locally in the northeast and southwest. Correlations with temperature are mainly positive (54.73%), but the proportion of significant pixels is relatively low (0.26% at p < 0.01 and 1.60% at p < 0.05), and spatially confined to the northeast, southeast and southwest.

Fig. 11
figure 11

The correlation between kNDVI in Inner Mongolia and average groundwater depth, precipitation and temperature. This base map is publicly available and has been included in the title of Fig. 1. This map was created by me using the open-source software QGIS 3.40.0.

Full size image

When considering both the correlation coefficients and significance levels, the strength of association with kNDVI follows the order: groundwater depth (− 0.97 to 0.95) > precipitation (− 0.85 to 0.94) > temperature (− 0.84 to 0.82). This indicates that groundwater availability exerts the strongest constraint on vegetation growth across Inner Mongolia, followed by precipitation inputs, whereas temperature plays a comparatively weaker role. Overall, water availability—jointly controlled by groundwater and precipitation—is the key driver of both spatial vegetation heterogeneity and interannual variability in this dryland region.

Ranking the explanatory power of soil–topography–LULC factors

Geodetector results show that a higher q value indicates stronger single-factor explanatory power. Among all examined variables, total nitrogen (TN) exhibits the highest explanatory strength (q = 0.41; Fig. 12), and areas with higher TN are generally associated with higher vegetation vitality. Soil organic matter (SOM), pH, and available phosphorus (AP) also show substantial contributions, confirming that higher SOM and AP and lower pH values (cf. Figures 13 and 4) tend to favour vegetation growth. Available nitrogen (AN), elevation (DEM), total potassium (TK), land use/land cover (LULC) and total phosphorus (TP) have moderate explanatory power, indicating that terrain characteristics, human land use, and selected nutrient levels act as important modulators of vegetation spatial patterns. In contrast, available potassium (AK), local relief, and slope contribute the least and have only marginal influence. Overall, spatial heterogeneity in vegetation growth across Inner Mongolia is mainly controlled by soil nutrient conditions—particularly TN, SOM and AP—while land use and topography act as background constraints that shape the regional expression of these nutrient effects.

Fig. 12
figure 12

Explanatory power of each influence factor.

Full size image
Fig. 13
figure 13

Spatial interpolation results of soil data. This base map is publicly available and has been included in the title of Fig. 1. This map was created by me using the open-source software QGIS 3.40.0.

Full size image

Discussion

kNDVI spatial and temporal characteristics

This study systematically assessed the spatiotemporal dynamics of kNDVI in Inner Mongolia from 2000 to 2024, and emphasized its advantages in arid and semi-arid ecosystems compared with the traditional NDVI. Unlike other studies relying on NDVI to monitor vegetation changes, kNDVI smooths NDVI time series using kernel density estimation and thus enhances its sensitivity to actual vegetation changes under mixed pixel conditions46,47,48. In areas characterized by low vegetation cover and transitional ecotones (e.g. agro‐pastoral ecotones and desert steppes), the present study found that kNDVI more effectively captured interannual variability and avoided the “noise amplification” effect of NDVI. By adjusting the σ value of the kernel function, kNDVI exhibits a higher stretching capacity for extreme values in time series, enabling the identification of fine‐scale vegetation responses under extreme climate conditions12. Importantly, we showed that kNDVI maintains a high dynamic sensitivity in regions with poor vegetation growth, which is consistent with its enhanced responsiveness to photosynthetic activities49,50,51,52. Moreover, kNDVI exhibited stable and consistent responses across different land‐use types (desert steppe, typical steppe and forest), suggesting a strong robustness of the index53. Integrating our field observation with temporal analysis, kNDVI clearly demonstrates its potential to detect subtle changes in the early stages of vegetation recovery, and therefore represents a valuable tool for evaluating ecosystem restoration in Inner Mongolia.

Applying this enhanced index, our analysis of Inner Mongolia from 2000 to 2024 revealed an overall increasing trend in kNDVI (Fig. 7b), indicating an overall improvement of the regional vegetation ecosystem driven by long-term ecological engineering and natural recovery processes54. However, a pronounced regime shift was detected around 2008, followed by a short period of decline and then a subsequent recovery. This pattern suggests that the positive effects of ecological engineering in the initial phase were offset by intensive human disturbances such as agricultural expansion and mining activities55,56. At the spatial scale, kNDVI exhibited a clear east–west gradient, with higher values in the forest–typical steppe region in the east and lower values in the desert and desert steppe regions of the west. This spatial heterogeneity is driven by regional differences in climate, soil properties and topography. From 2008 to 2016, continuous declines were observed in Baotou and Siziwang Banner, representing a typical case of combined effects of policy implementation, drought and overgrazing57,58,59. It is noteworthy that after 2016, kNDVI in the Alxa League (western region) sharply decreased and the projection results of this study suggest a further degradation trend in the near future60. This indicates that vegetation recovery may cross a critical threshold once rainfall scarcity and groundwater depletion coexist. Consequently, areas that show early signs of reversed vegetation improvement need continuous monitoring and field verification to prevent a transition towards irreversible degradation61,62,63.

Policy implication

The increase in kNDVI between 2000 and 2024 reflects the combined effects of natural conditions and policy-driven interventions. In arid and semi-arid regions, groundwater availability is a critical factor for sustaining vegetation growth64. In the desert and sandy areas of Inner Mongolia, the significant negative correlation between groundwater depth and kNDVI suggests that declining groundwater tables intensify water stress and inhibit vegetation recovery. Conversely, the positive relationship observed in the Hulunbuir grasslands demonstrates that deep-rooted species can effectively utilize deeper groundwater resources65, underlining the importance of conserving groundwater as part of restoration initiatives.

From a climatic perspective, precipitation is the primary driver of vegetation change in grasslands and cropland areas, while temperature is more influential in forest regions66. However, the increasing frequency of extreme events such as heat waves and droughts has imposed additional pressure on vegetation, particularly in the western part of the region where kNDVI shows rapid declines and slow post-event recovery67,68. These patterns highlight the need to incorporate climate risk into regional ecological management and to enhance resilience against extreme events.At the same time, soil quality and topographic features exert indirect yet crucial effects by regulating soil water retention, nutrient supply and energy balance, thereby contributing to the spatial heterogeneity of vegetation responses69,70. Importantly, land-use policies such as reforestation and grazing exclusion have clearly promoted vegetation recovery, whereas areas subject to continued intensive use show weaker kNDVI improvement. This confirms that policy interventions play a decisive role in regulating vegetation dynamics, and that effective restoration in drylands must integrate hydrological management (e.g., groundwater conservation), climate adaptation measures and sustainable land-use planning.

Limitations and future perspectives

Although this study demonstrates the effectiveness of kNDVI for monitoring vegetation and reveals clear spatial patterns of recovery and degradation across Inner Mongolia, several limitations should be acknowledged. The use of 30 m Landsat imagery is appropriate for regional analyses but may overlook patch-scale changes, particularly in highly heterogeneous terrain where localized degradation or early-stage recovery processes can be obscured. Integrating higher-resolution remote sensing products (e.g., Sentinel, GF, PlanetScope) and adopting multi-scale data fusion techniques would therefore improve the spatial sensitivity of vegetation monitoring71.

In addition, Since land use types cannot be quantified and topographic factors and soil chemical properties are single-year data, the pixel-wise correlations method cannot be used to obtain their spatial correlations with kNDVI. Therefore, geodetectors are used to obtain their influence on kNDVI. vegetation dynamics in dryland systems are influenced not only by climate and groundwater conditions but also by ecological processes such as evapotranspiration, surface runoff and wind erosion72. These dynamic processes were not explicitly incorporated into the present analysis and may represent important drivers in specific geomorphic units. The inclusion of surface process models and advanced machine learning approaches would help capture these interactions and allow for a more comprehensive attribution of vegetation change, ultimately improving the predictive capacity and management relevance of future studies.

Conclusions

Based on the kNDVI index, this study provided a comprehensive assessment of vegetation dynamics and their drivers across Inner Mongolia from 2000 to 2024. Compared with the traditional NDVI, the kNDVI index showed a much higher sensitivity in detecting subtle vegetation changes, particularly in low-cover and ecologically fragile areas. This feature is especially important for monitoring early-stage ecosystem recovery in arid and semi-arid regions. The temporal analysis revealed a general improvement in vegetation growth over the past 25 years, but this improvement occurred in a clearly stage-dependent manner. The abrupt change around 2008 indicates the presence of critical thresholds and highlights the vulnerability of vegetation recovery to human disturbances. From a spatial perspective, vegetation restoration was most pronounced in the eastern part of the region, whereas the western desert and parts of the central–eastern zone still exhibited noticeable degradation and remain at risk of continued decline. The driving factor analysis further shows that, in contrast with the often-emphasized temperature and precipitation controls, groundwater depth exerts the strongest influence on vegetation dynamics in Inner Mongolia. This effect is particularly evident in the desert and sandy ecosystems, where falling groundwater levels significantly exacerbate water stress. In addition, Pearson correlation and Geodetector results demonstrate that soil nutrient conditions – especially total nitrogen – play a decisive role in shaping the spatial pattern of vegetation growth. These findings indicate that successful ecosystem restoration in arid regions requires an integrated perspective that simultaneously considers vegetation–groundwater coupling and soil quality improvement, rather than relying solely on precipitation increases or afforestation/closure measures.