Integrating Multi-Temporal Sentinel-1/2 Vegetation Signatures with Machine Learning for Enhanced Soil Salinity Mapping Accuracy in Coastal Irrigation Zones: A Case Study of the Yellow River Delta

Abstract

Soil salinization poses a severe threat to agricultural sustainability in the Yellow River Delta, where conventional spectral indices are limited by vegetation interference and seasonal dynamics in coastal saline-alkali landscapes. To address this, we developed an inversion framework integrating spectral indices and vegetation temporal features, combining multi-temporal Sentinel-2 optical data (January 2024–March 2025), Sentinel-1 SAR data, and terrain covariates. The framework employs Savitzky–Golay (SG) filtering to extract vegetation temporal indices—including NDVI temporal extremum and principal component features, capturing salt stress response mechanisms beyond single-temporal spectral indices. Based on 119 field samples and Variable Importance in Projection (VIP) feature selection, three ensemble models (XGBoost, CatBoost, LightGBM) were constructed under two strategies: single spectral features versus fused spectral and vegetation temporal features. The key results demonstrate the following: (1) The LightGBM model with fused features achieved optimal validation accuracy (R² = 0.77, RMSE = 0.26 g/kg), outperforming single-feature models by 13% in R². (2) SHAP analysis identified vegetation-related factors as key predictors, revealing a negative correlation between peak biomass and salinity accumulation, and the summer crop growth process affects soil salinization in the following spring. (3) The fused strategy reduced overestimation in low-salinity zones, enhanced model robustness, and significantly improved spatial gradient continuity. This study confirms that vegetation phenological features effectively mitigate agricultural interference (e.g., tillage-induced signal noise) and achieve high-resolution salinity mapping in areas where traditional spectral indices fail. The multi-temporal integration framework provides a replicable methodology for monitoring coastal salinization under complex land cover conditions.

1. Introduction

Soil salinization is globally recognized as a major driver of land degradation, posing a serious threat to agricultural productivity, ecosystem integrity, and food security. The excessive accumulation of soluble salts in the soil profile not only inhibits plant growth and reduces crop yields but ultimately leads to the abandonment of affected agricultural land. These changes impair critical ecosystem services and accelerate the process of desertification [1,2]. Salt accumulation is exacerbated by high evaporation rates and inadequate drainage systems, while unsustainable irrigation practices and extensive land management are identified as key anthropogenic drivers worldwide [1,3,4]. Soil structure and fertility are degraded by salinization, while hydrological cycles are disrupted, biodiversity is reduced, and the resilience of both natural and managed ecosystems is significantly impaired [2,5,6]. The resulting socioeconomic impacts are profound, jeopardizing the livelihoods of millions and increasing the vulnerability of communities to the effects of climate change and water scarcity [1,3,5].

The Yellow River Delta is recognized as both a critical agricultural production zone and an important wetland reserve. However, the region is confronted with severe coastal salinization, driven by a high groundwater table, frequent seawater intrusion, and limited freshwater recharge. It has been demonstrated that 70–76% of the area is affected by saline-alkali soils [7,8,9,10]. Under the influence of intense land–sea interactions, particularly seawater intrusion, and agricultural irrigation activities, the soil structure is degraded, fertility is reduced, and plant growth is inhibited by salinization, subsequently leading to crop yield reduction and the loss of arable land [8,11,12]. Widespread soil salinization has been identified as a major constraint to local sustainable development.

However, traditional ground surveys of soil salinization are constrained by high costs, time-consuming procedures, and limited spatial coverage, making large-scale, high-frequency monitoring challenging. Satellite remote sensing technology has been recognized as the only feasible technical means for monitoring soil salinity at a regional scale due to its efficiency, objectivity, and capacity for periodic data acquisition [13,14,15]. Through the establishment of mathematical models that correlate remote sensing data with regional soil salinity parameters, satellite-based inversion techniques enable the dynamic simulation of the spatial distribution of soil salinity and provide accurate, timely spatial quantitative information [13,14,15]. The Multispectral Imager (MSI) on the European Space Agency’s (ESA) Sentinel-2 satellite provides spectral data from the visible to the shortwave infrared range, which can be utilized to derive spectral indices related to vegetation health and surface soil properties. As a complementary approach, the C-band Synthetic Aperture Radar (SAR) on Sentinel-1 is not limited by illumination or weather conditions, and the backscatter intensity and coherence metrics provided by it are sensitive to soil moisture, surface roughness, and dielectric properties—parameters that are indirectly related to salt content [16,17,18]. Consequently, this synergistic use of sensors provides a robust data foundation for satellite-based soil salinity monitoring [16,17,18]. Based on the characteristic spectral responses of saline soils, various salinity indices have been developed by researchers. However, the diagnostic sensitivity of these indices in practical application scenarios is significantly reduced by confounding factors, including vegetation canopy cover and soil moisture content. Therefore, such indices are primarily applicable in arid and semi-arid regions [16,19]. Research has indicated that crop coverage and growth stages regulate the distribution of soil salinity—salinity is generally lower during peak growth periods (summer) and higher in deep or bare soil layers, with optimal monitoring accuracy typically achieved in bare soil or during early crop stages. In irrigated areas, as crop coverage increases over time, salinity in the root zone can be reduced by crop transpiration and irrigation practices (e.g., drip irrigation), which may lead to a decrease in inversion accuracy [20]. However, this trend may be slowed or reversed during periods of reduced irrigation or toward the end of the growing season [21]. In irrigated regions, crop activities significantly influence salinity dynamics and the performance of inversion models. Agricultural practices—including irrigation methods, tillage, crop management, and residue handling—have a direct impact on both the accumulation of soil salinity and the accuracy of remote sensing-based salinity inversion. These practices affect the movement, distribution, and spatial patterns of salinity within agricultural landscapes. Another critical aspect is temporal dynamics: seasonal variations in agricultural activities (e.g., planting, irrigation, harvesting) cause temporal fluctuations in salinity distribution [22,23]. Therefore, the influence of vegetation dynamics must be accounted for in remote sensing models.

Research has shown that the integration of spectral, topographic, and ecological variables as covariates can improve salinity assessment, though the lack of multi-temporal data remains a key bottleneck [24,25,26]. Studies in multiple regions, including the Yellow River Delta, Murray–Darling Basin, and Mekong Delta, have confirmed that while vegetation indices are valuable, their performance is context-dependent and can be greatly enhanced through the fusion of multi-temporal and multi-source data [27,28]. Time-series features, especially those derived from vegetation indices, serve as crucial indicators by capturing cumulative or persistent vegetation stress patterns driven by underlying salinity levels. Overall, the use of multi-temporal vegetation indices and advanced modeling techniques has been demonstrated to address many shortcomings of traditional single-date approaches, enabling more accurate and comprehensive soil salinity monitoring across diverse landscapes and salinity gradients [29]. However, under conditions of limited sample size, the establishment of robust relationships between remote sensing data and soil salinity predictors is particularly challenging, as linear models are often unable to capture the complex non-linear interactions present in diverse environments [30,31]. Machine learning (ML) techniques, including Random Forest, XGBoost, Support Vector Machines (SVM), and deep learning architectures, have been demonstrated to outperform traditional linear models by effectively simulating complex relationships even with small datasets [32,33,34]. These methods have been successfully applied to predict and map soil salinity using various remote sensing inputs, such as spectral indices from Sentinel-2, microwave backscatter from Sentinel-1, and time-series vegetation indices, with these features often being integrated into unified models to improve accuracy. Studies have consistently shown that ML models surpass linear regression in both accuracy and generalization ability, while ensemble and hybrid approaches (e.g., combining deep learning with tree-based models) further enhance predictive performance [35]. Through the integration of multi-source satellite and ground data with machine learning (e.g., Random Forest, SVM, deep learning), along with feature selection strategies, robust inversion of soil salinity has been achieved for both bare and vegetated soil. Land cover can be classified by ML models, enabling tailored inversion strategies that ensure reliable performance across different surface conditions [34,35]. Nevertheless, for modeling strategies focusing on spectral and vegetation time-series, the influence of feature optimization within the modeling process remains unclear in coastal agricultural irrigation districts.

This study proposes the optimal vegetation type signature (OVTS) method. OVTS is defined as a set of temporal features (including extremum, mean, and principal compo-nent features, among others) extracted from the Savitzky-Golay filtered NDVI time series, which best represent vegetation types and their dynamic responses to salt stress. The approach aims to elucidate the impact of vegetation dynamics on salt accumulation by synergistically integrating Sentinel-1 SAR data and multi-temporal Sentinel-2 sequences. Compared to traditional optical spectral indices (OS), OVTS offers a dual advantage: 1) its multi-temporal feature set captures the dynamic effects of salt stress over time, revealing the influence of vegetation temporal characteristics on salt formation processes; and 2) vegetation temporal features reduce uncertainties caused by disturbances from agricultural activities (e.g., spring tillage and crop residue incorporation), thereby enhancing the robustness of soil salinity inversion. This study aims to address two key research questions: 1) Can the OVTS-enhanced strategy strengthen the relationship between vegetation temporal characteristics and soil salinity to improve salt monitoring accuracy? 2) Identify the optimal temporal features for OVTS-based salinity inversion to achieve high-precision soil salinity mapping in the Yellow River Irrigation District.

2. Materials and Methods

The research framework implemented in this study was structured around three key procedures: (1) remote sensing data preprocessing; (2) modeling strategies based on OS and OVTS; (3) accuracy evaluation and results analysis. Figure 1 illustrates the research framework flowchart, while comprehensive technical specifications for each methodological phase will be detailed in subsequent sections.

2.1. Study Area

The study area is situated in the center of the Yellow River Delta, adjacent to the Bohai Sea (Figure 2). The Yellow River irrigation area serves as a critical ecological reserve, with young and porous soils, and significant seawater intrusion. Elevated groundwater tables and increased total dissolved solids (TDS), driven by seawater intrusion and reduced river discharge, constitute key drivers of salinity expansion [7,36]. These environmental factors render the region highly susceptible to soil salinization and alkalization, triggering extensive land degradation and ecological deterioration. The climate is warm-temperate continental with a mean annual precipitation of 576.7 mm, a mean annual potential evapotranspiration of 1962 mm, a mean annual temperature of 12.1 °C, and the frost-free period 196 d. Dominant soil types include fluvo-aquic soil and coastal saline soil, with the former being closely associated with alluvial deposits derived from historical sedimentation processes of the Yellow River. Proximity to the ocean, with a high groundwater level, and seawater intrusion contribute significantly to elevated soil salinity in the region. The pattern of vegetation distribution follows soil salinity gradients: highly saline zones are dominated by native halophytic species such as Suaeda salsa, Tamarix chinensis, and Phragmites australis, whereas major crops including winter wheat, maize, cotton, and soybean are cultivated in inland areas with lower salinity [37]. Spatially, salinization intensifies progressively from the southwestern inland to the northeastern coast, with coastal zones and ancient river channels exhibiting persistent aggravation [10,38]. The ongoing interplay of climatic, hydrological, and land use factors continues to shape the delta’s salinity gradient landscape, making precise salinization monitoring and adaptive management of saline farmlands essential for ecological conservation and sustainable agricultural development.

2.2. Sample Collection and Laboratory Measurements

Crucial field sampling targeting the initial soil surface salinity accumulation phase was conducted in the coastal region of the northeastern Yellow River Delta from 18 March to 23, 2025, to capture peak salinity levels critical for evaluating ecological management and agricultural practices [39]. Sampling sites were strategically distributed, integrating considerations of vegetation type, vegetation coverage, land use type, and known spatial variation patterns of soil salinity, thereby effectively capturing spatial heterogeneity to establish a robust foundation for precision mapping and model calibration. Soil samples were collected using a standardized five-point sampling method within 10 m × 10 m quadrats from the 0–20 cm depth, yielding a total of 119 composite samples. The Global Positioning System (GPS) coordinates of each sampling point and key environmental parameters (e.g., surface cover, surrounding environmental features) were recorded. Following collection, soil samples were air-dried at ambient temperature in shaded conditions, ground, and sieved through a 2 mm mesh. For salinity determination, 20 g of the processed soil was mixed with 100 mL of distilled water at a 1:5 (w/v) soil-to-water ratio to prepare a 1:5 soil-water extract, the electrical conductivity (EC₁:₅) of which was subsequently measured by conductivity meter of DOS-307A. Total soil salinity content was calculated using the ion summation method by ICP-AES Avio 550max.

2.3. Acquisition of Environment Variables

2.3.1. Remote Sensing Data

To generate a soil salinity map for the Yellow River Irrigation Zone, this study utilized Sentinel-2/Sentinel-1 satellite remote sensing data. Specifically, 10 m resolution imagery acquired from January 2024 to March 2025 (sourced from the Google Earth Engine dataset catalog: COPERNICUS_S2_SR_HARMONIZED). The Google Earth Engine (GEE) platform provides access to extensive earth observation datasets and leverages cloud computing technology for the efficient processing of petabyte-scale satellite imagery, enabling robust large-scale analysis of the study area. Consequently, Sentinel-2 multispectral imagery within the GEE environment was employed to calculate several commonly used predictive variables for soil salinity inversion. Spectral indices at the image acquisition date and vegetation time series characteristics were derived based on Sentinel-2 data processed on GEE. Soil salinity inversion is not applicable in areas classified as construction land. Therefore, building pixels were excluded using the European Space Agency (ESA) WorldCover 10 m 2020 land cover product.

2.3.2. Acquisition of Spectral Indices from Sentinel-2

Soil salinity is influenced by the integrated effects of soil, climate, topography, and biology across different scales. Therefore, in digital soil mapping, the selection of constructed indices must respond to these soil-forming factors. To build a streamlined yet informative feature set for predicting soil salinity, a series of validated spectral indices are calculated in this study. These indices are primarily categorized into three groups: (1) general vegetation indices, which are used to indirectly assess changes in vegetation vigor caused by salt stress; (2) salinity-specific indices, which utilize spectral features sensitive to the physicochemical properties of the soil surface; and (3) tasseled cap components, which are derived to capture macro-scale surface changes related to salinity. Detailed descriptions of these indices are provided in

Table 1. Spectral predictors: abbreviations, calculation formulas, and references.

Catalog	Predictors	Abbreviations	Formulations	Reference
Original Bands	Blue	B	/	Sentinel-2
	Green	G	/	Sentinel-2
	Red	R	/	Sentinel-2
	Red Edge1	VRE1	/	Sentinel-2
	Red Edge2	VRE2	/	Sentinel-2
	Red Edge3	VRE3	/	Sentinel-2
	Nir	NIR	/	Sentinel-2
	Red Edge4	VRE4	/	Sentinel-2
	SWIR1	SWIR1	/	Sentinel-2
	SWIR2	SWIR2	/	Sentinel-2
Soil property indices	Clay Index	CLEX	SWIR1/SWIR2	(Taghizadeh-Mehrjardi et al., 2014) [41]
	Carbonate Index	GAEX	G/B	(Taghizadeh-Mehrjardi et al., 2014) [41]
	Gypsum Index	GYEX	(SWIR1 − NIR)/(SWIR2 + NIR)	(Taghizadeh-Mehrjardi et al., 2014) [41]
Vegetation indices	Extended EVI	EEVI	2.5 * [(NIR + SWIR2-R)/(NIR + SWIR2 + 6 * R − 7.5 * B + 1)]	(Ma et al., 2023) [42]
	Soil Adjusted Vegetation Index	SAVI	[(NIR − R) * 1.5]/(NIR + R + 0.5)	(Huete, 1988) [43]
	Extended NDVI	ENDVI	(NIR + SWIR2-R)/(NIR + SWIR2 + R) −	(Chen et al., 2015) [44]
	Generalized Difference Vegetation Index	GDVI	(NIR2 − R2)/(NIR2 + R2)	(Wu et al., 2014) [45]
	Global Vegetation Moisture Index	GVMI	[(NIR + 0.1) − (SWIR1 +0.02)]/[(NIR + 0.1) + (SWIR1 + 0.02)]	(Ceccato et al., 2002) [46]
	Infrared Percentage Vegetation Index	IPVI	NIR/(NIR + R)	(Crippen, 1990) [47]
	Normalized difference vegetation index	NDVI	(NIR − R)/(NIR + R)	(Rouse Jr et al., 1974) [48]
	Normalized difference water index	NDWI	(G − NIR)/(G + NIR)	(McFeeters, 1996) [49]
	Enhanced Residues Soil Salinity Index	ERSSI	G2/B * SWIR1	(Wang, et al., 2022) [50]
Salinity indices	Brightness Index	BI	(G2 + B2)0.5	(Khan et al., 2005) [51]
	Salinity index I	S1	B/R	(Khan et al., 2005) [51]
	Salinity index II	S2	(B − R)/(B + R)	(Khan et al., 2005) [51]
	Salinity index III	S3	G * R/B	(Khan et al., 2005) [51]
	Salinity index V	S5	B * R/G	(Khan et al., 2005) [51]
	Salinity index VI	S6	R * NIR/G	(Khan et al., 2005) [51]
	Salinity Index 1	SI1	(G + R)0.5	(Khan et al., 2005) [51]
	Salinity Index 2	SI2	(NIR2 + G2 + R2)0.5	(Khan et al., 2005) [51]
	Salinity Index 3	SI3	(G2 + R2)0.5	(Khan et al., 2005) [51]
	Salinity Index 4	SI4	SWIR1/NIR	(Douaoui et al., 2006) [52]
	Canopy Response Salinity Index	CRSI	[(NIR * R − G * B)/(NIR * R + G * B)]0.5	(Scudiero et al., 2014) [53]
	Normalized Difference Salinity Index	NDSI	(NIR − SWIR1)/(NIR + SWIR1)	(Major et al., 2007) [54]
	Salinization Remote Sensing Index	SRSI	[(NDVI − 1)2 + SI2]0.5	(Alhammadi and Glenn, 2008) [55]
	Salinity index VII	S7	(SWIR1 − SWIR2)/(SWIR1 + SWIR2)	(Bannari et al., 2008) [56]
	Soil Salinity and Sodicity Indices1	SSS_1	R-NIR	(Bannari et al., 2008) [56]
	Soil Salinity and Sodicity Indices2	SSS_2	(R * NIR − NIR2)/R	(Bannari et al., 2008) [56]
Radar indices	Backscattering coefficients of VH band	VH	(ơ0 − ơ0veg_VH)/L	(Kumar et al., 2012) [57]
	Backscattering coefficients of VV band	VV	(ơ0 − ơ0veg_VV)/L	(Kumar et al., 2012) [57]
TC	Tasseled cap transformation of Sentinel-2 bands	TC1,TC2,TC3	/	/
OVTS	Principal components of NDVI time-series	NDVI_PC1, NDVI_PC2, NDVI_PC3	/	/
	Maximum, minimum, and mean of NDVI time-series	NDVI_max, NDVI_min, NDVI_mean	/	/
Terrain indices	Derivative Topographic Metrics	Aspect, Elevation, Slope	/	/

Note: The asterisk symbol * denotes the logical multiplication operator in all equations.

. Vegetation indices are often employed as key parameters for indirectly indicating vegetation growth conditions under soil salinity stress. To directly capture soil salinity information, several established salinity indices are introduced. The Salinity Index (SI) utilizes ratio relationships in the visible spectrum and shows a significant response to salt crusts or efflorescence on the soil surface, having been widely used for mapping saline soils in arid and semi-arid regions [35]. The three core components of the tasseled cap transformation—Brightness (TC1), Greenness (TC2), and Wetness (TC3)—are derived from Sentinel-2 imagery. TC1 (Brightness) is positively correlated with the degree of surface exposure and soil brightness, and high salinity areas often exhibit high brightness values due to sparse vegetation and dry, salt-rich soil surfaces. TC2 (Greenness) is closely associated with vegetation coverage and can be used to monitor vegetation degradation induced by salt stress. TC3 (Wetness) is sensitive to the moisture content of both soil and vegetation, and since salinity affects soil moisture conditions, TC3 can serve as an indirect indicator of salinity. Multiple studies have successfully applied tasseled cap components to regional soil salinity monitoring and assessment [40].

2.3.3. Vegetation Temporal Metrics Derived from Sentinel-2

The Normalized Difference Vegetation Index (NDVI) is one of the most widely used vegetation indices and serves as the foundation for constructing OVTS in remote sensing studies [58]. In this research, Sentinel-2 imagery from January 2024 to March 2025 was processed to generate continuous, gap-free NDVI time series for each pixel across the study area, with the Savitzky-Golay filter (16-day window, first-order polynomial) applied to reduce residual noise and enhance signal quality [59]. The SG filter is recognized for its effectiveness in smoothing NDVI time series and preserving phenological patterns, outperforming simpler methods like moving averages. Distinct NDVI temporal profiles were observed for different vegetation types: winter wheat and summer maize exhibited a characteristic double-peak pattern reflecting a double-cropping system, while halophytic vegetation showed lower maximum NDVI values. Principal Component Analysis (PCA) was then applied to the smoothed NDVI time series, extracting the main modes of variation and generating principal components (PCS) images as NDVI_PC1, NDVI_PC2, and NDVI_PC3, which capture the dominant shape and amplitude dynamics of the NDVI curves. PCA is particularly effective for land cover mapping, seasonal change detection, and analysis of phenological drivers, as it transforms correlated time series data into orthogonal components that summarize key temporal features [60]. To further characterize vegetation activity, statistical metrics such as maximum, minimum, and mean NDVI were calculated, providing additional information on the amplitude and variability in the time series. This integrated approach combining NDVI time series, SG filtering and PCA have been widely validated in regional scale vegetation and land use mapping, crop phenology analysis, and environmental monitoring [61].

2.3.4. Sentinel-1 and DEM Predictor Variables

Sentinel-1 Ground Range Detected (GRD) imagery was acquired from the ESA Copernicus Open Access Hub and subjected to standard preprocessing workflows. This included orbit correction using precise orbital ephemerides, removal of thermal noise through Sigma0 calibration, and radiometric calibration to derive backscatter coefficients (σ⁰). The processed data covered the entire study area during the field sampling campaign (18–23 March 2025). Calibrated backscatter intensities for both cross-polarization (VH: Vertical Transmit, Horizontal Receive) and co-polarization (VV: Vertical Transmit, Vertical Receive) channels were extracted as salinity-sensitive indicators, leveraging the established relationship between salinization-induced alterations in soil dielectric properties and microwave scattering mechanisms. For topographic parameterization, the NASA Shuttle Radar Topography Mission (SRTM) Global 30 m DEM (GL1 v3) was resampled to 10 m spatial resolution using bilinear interpolation to maintain consistency with the optical satellite data grid, enabling subsequent derivation of terrain covariates.

2.4. Modeling

2.4.1. Variable Selection

VIP (Variable projection importance) is a statistically robust feature selection method widely used in Partial Least Squares Regression (PLSR) to identify predictors with significant explanatory power for the response variable. VIP quantifies each predictor’s overall contribution to modeling the response by evaluating its weighted influence across all PLSR components, providing a comprehensive measure of importance that accounts for the variance explained by each latent variable. Before modeling, all predictors are standardized to ensure comparability. The optimal number of PLSR components is typically determined via cross validation commonly 10-fold cross validation to minimize overfitting and enhance model generalizability. VIP-based selection is especially valuable in high dimensional or collinear datasets, as it reduces dimensionality while maintaining predictive accuracy and interpretability. Since many constructed indices in Section 2.3.2 were significantly correlated with measured salinity, the VIP method was employed to screen these remote sensing indices to simplify the input variables for the salinity inversion model. The indices with VIP values greater than 1 were selected to serve as input variables for salinity inversion. The VIP selection was implemented using Python (version 3.6.0) packages: numpy, pandas, and scikit-learn. The formulas for these VIP metrics are presented below:

$${\mathrm{V}\mathrm{I}\mathrm{P}}_k=\sqrt{{\displaystyle \frac{p}{{\sum }_{h=1}^H {\mathrm{S}\mathrm{S}\mathrm{Y}}_h}}{\sum }_{h=1}^H {\mathrm{S}\mathrm{S}\mathrm{Y}}_h{\left(w_{kh}/‖{\mathbf{w}}_h‖\right)}^2}$$

(1)

where p represents the total number of predictors; H represents the number of retained PLSR components; SSY_h represents the explained variance in Y by the h-th component; w_kh represents the PLSR weight of the k-th predictor for component h; and ‖w_h‖ represents the Euclidean norm of the weight vector for component h.

2.4.2. Modeling Method and Evaluation Indicators

Three advanced ensemble algorithms based on the Gradient Boosting Decision Tree (GBDT) framework—XGBoost, CatBoost, and LightGBM—were employed for soil salinity prediction in this study, a methodology that has been widely applied in digital soil mapping and environmental modeling [62,63]. All models were implemented using Python packages from the Scikit-learn library within the Google Earth Engine (GEE) Colab environment. Through an optimal parameter grid search strategy, the detailed configurations of each model are presented in

Table 2. Optimal model hyperparameter combination.

Models	N_Estimators	Learning Rate	Max Depth	Reg_Alpha	Reg_Lambda	Subsample
XGBoost	100	0.1	3	2	3	0.6
CatBoost	100	0.1	6	/	3	0.6
LightGBM	100	0.05	10	1	1	0.6

. To establish a rigorous model evaluation framework and mitigate overfitting risks, 10-fold cross-validation was applied to all models, with stratification based on quantiles of measured soil salinity values to ensure consistency in data distribution. Model performance was evaluated using the Root Mean Square Error (RMSE) and the coefficient of determination (R²). The mean RMSE and R² values, which were aggregated from all validation folds, provided a comprehensive assessment of the predictive performance and generalization capability of each algorithm [62]. The mathematical formulations of RMSE and R² are given in Equations (2) and (3), respectively.

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

(2)

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

(3)

where ${\widehat{y}}_i$ and $y_i$ represent the model estimates and the true values, respectively; $\bar{y}$ denotes the mean of the true values; and n is the number of data points.

2.5. SHAP Analysis and Feature Importance

To elucidate the decision-making mechanisms of models under multi-feature synergistic effects, this study introduces the (Shapley Additive exPlanations) SHAP interpretability framework based on game theory. This approach conceptualizes machine learning model outputs as cooperative games among features, quantifying the marginal contribution of each feature to predictive outcomes through Shapley value computation. This enables investigation into the influence of features on model predictions during training. The computational process adheres to the following axiomatic principles: SHAP efficiently resolves Shapley values in high-dimensional feature spaces via kernel approximation (Kernel SHAP) or tree structure-optimized algorithms (Tree SHAP). Compared with traditional feature importance assessment methods, SHAP provides enhanced interpretability by assigning importance values to each feature, thereby revealing the decision pathways underlying predictive outcomes.

3. Results

3.1. Descriptive Statistics and Spatial Autocorrelation Analysis

Based on the statistical analyses of the range, standard deviation (SD), coefficient of variation (CV), and distribution characteristics of the soil salinity dataset from all 119 sampling points in

Table 3. Descriptive statistics of the soil dataset for the total, modeling and validation set (All relevant units are in g/kg).

Data	N	Max	Min	Mean	STD	CV	Skewness
Total data	119	16.49	0.45	3.11	3.61	1.16	2.01
Modeling data	95	14.16	0.45	2.69	2.96	1.1	2.06
Verification data	24	16.49	0.71	4.77	5.22	1.09	1.27

, the soil data in the study area exhibit significant variability. The overall sample distribution demonstrates strong variability and a right-skewed pattern: the entire dataset has a CV of 1.16, with salinity values ranging from 0.45 to 16.49 g/kg and a skewness value of 2.01, indicating a right-skewed distribution. This suggests a highly discrete spatial distribution of soil salinity and a distinct salinization gradient. The dataset was partitioned into a modeling set (80%) and a validation set (20%), with SDs of 2.96 g/kg and 5.22 g/kg, corresponding CVs of 1.10 and 1.06, and skewness values of 2.06 and 1.27, respectively, both indicating right-skewed distributions.

Given this substantial variability (CV > 1), logarithmic transformation was applied to the target variable to meet normality requirements. Based on the spatial autocorrelation analysis using Moran’s I index, the spatial distribution pattern and statistical significance of soil salinity were evaluated. The results demonstrated that soil salinity exhibited a significant spatially clustered pattern (Figure 3). The Moran scatter plot clearly illustrated the spatial association characteristics of soil salinity. The majority of sampling points were distributed in the first quadrant (high-high clustering) and the third quadrant (low-low clustering), indicating a pronounced positive spatial autocorrelation in soil salinity distribution and revealing a distinct spatially aggregated pattern. Specifically, the High-High (HH) clustering areas may represent core zones of coastal soil salinization, while the Low-Low (LL) clusters indicate regions with relatively better soil quality, typically associated with irrigated agriculture. In contrast, the Low-High (LH) and High-Low (HL) clusters likely reflect transitional zones or outliers in salinity distribution, whose anomalous patterns may arise from localized environmental factors or anthropogenic disturbances. The test results further confirmed the statistical significance of spatial autocorrelation. The Moran’s I index value of 0.22 (p < 0.001) was significantly greater than the expected value under random distribution, indicating a moderately strong yet highly significant positive spatial autocorrelation in soil salinity. The Z-score reached 6.64, far exceeding the critical value at the 0.05 significance level, demonstrating that this spatial pattern did not occur by chance but rather represented a spatially structured characteristic with high statistical significance. This pronounced spatial autocorrelation may originate from the combined effects of multiple environmental factors, including spatially autocorrelated variables such as groundwater depth, soil texture, irrigation practices, and land use types.

3.2. Environment Variable Selection

Figure 4 presents a radar chart illustrating the distribution characteristics of Pearson correlation coefficients between 52 covariates derived from Sentinel-2 multispectral data, Sentinel-1 radar data, and DEM data with measured soil salinity. The feature variables are categorized into six distinct classes: Original Bands (OB), Vegetation Index Series (VIS), Vegetation Indices (VI), Salinity Indices (SI), Soil Property Indices (SPI), and Radar/Terrain Indices (RD/Terrain). The results demonstrate that 35 multispectral covariates exhibit statistically significant correlations with soil salinity content (p < 0.05). Vegetation temporal features show the strongest negative correlation, while parameters including VV-polarized radar data, S4, B6, NDVI_PC2, tasseled cap wetness component (TC3), and aspect show no significant correlations. Surface moisture indices along with blue, green, and red bands (B2, B3, B4) also show significant correlations. Notably, traditional salinity indices generally exhibit weak correlations with measured salinity in coastal saline-alkali soils, confirming that spectral salinity indices are more suitable for arid/semi-arid regions. The findings from Figure 4 provide critical basis for feature selection, with vegetation temporal features demonstrating the highest predictive potential. Based on this analysis, two modeling strategies were established using the Variable Importance in Projection (VIP) method: one based on optimal spectral index features, and the other combining spectral indices with vegetation temporal features, thereby effectively addressing the dimensionality explosion issue in high-dimensional feature spaces. The detailed composition of characteristic variables included in two strategies is presented in

Table 4. The composition of feature factors selected based on VIP for the two strategies.

Scheme 1.	Indices Combination
Optical Signatures (OS)	CRSI,S1,S2,S7,SI2,DEM,B2, B3,B8,B12,BI,CLEX,ERSSI,EEVI
Optical Vegetation Type Signatures (OVTS)	NDVI_PC1,NDVI_max,NDVI_mean,CRSI,S1,S2,S7,SI2,DEM,B2,B3,BI,ERSSI,EEVI

.

3.3. Estimation of Soil Salinity Content Under Different Model Strategies

Soil salinity inversion models were established using XGBoost, CatBoost, and LightGBM algorithms, following feature selection based on the VIP method to identify significant variables. Notably, the three models utilizing spectral indices alone demonstrated lower accuracy than those incorporating fused spectral and temporal features. Comparative models employing only single-temporal spectral features (excluding OVTS) exhibited significant degradation in validation accuracy: XGBoost (R² decreased from 0.70 to 0.64), CatBoost (R² decreased from 0.68 to 0.58), and LightGBM (R² decreased from 0.77 to 0.68). The LightGBM model integrating OS and OVTS features achieved optimal performance, yielding a calibration R² = 0.92 and RMSE = 0.06. Its validation accuracy surpassed that of other models (R² = 0.77, RMSE = 0.26) (

Table 5. The model accuracy with spectral and temporal vegetation features.

Datasets	Model	Calibration		Validation
		R2	RMSE	R2	RMSE
Optical Signatures (OS)	XGBoost	0.80	0.39	0.64	0.63
	CatBoost	0.74	0.45	0.58	0.69
	LightGBM	0.89	0.08	0.68	0.36
Optical Vegetation Type Signatures (OVTS)	XGBoost	0.82	0.37	0.70	0.57
	CatBoost	0.81	0.38	0.68	0.60
	LightGBM	0.92	0.067	0.77	0.26

). Figure 5a–f scatter plots visually demonstrate the fitting patterns between predicted and measured values for the OS and OVTS models. Predictions were overestimated in low-salinity zones, with model predictions consistently exceeding measured values, particularly in the OS model (Figure 5a–c). Predictions were underestimated in high-salinity zones, which was partially mitigated but still present in the OVTS model [35]. This discrepancy may be attributed to vegetation interference: high spring vegetation coverage in the Yellow River Delta obscures bare soil spectral information, masking salinity signals beneath vegetation. During spring plowing, deep tillage disturbs surface soil, and residual crop stubble alters soil reflectance characteristics, leading to concurrent low NDVI and low salinity values (e.g., NDVI < 0.2). Overall, LightGBM model demonstrated enhanced robustness, attributable to the incorporation of vegetation temporal features (Figure 5f). Collectively, the integration of multi-temporal indices (OVTS) substantially improved model accuracy and generalization capacity, conclusively demonstrating that coupling OVTS characteristics with single-temporal remote sensing imagery effectively enhances soil salinity inversion capability.

3.4. SHAP Value and Feature Importance Analysis

Figure 6 reveals the feature importance of the optimal modeling strategy and the SHAP value results. SHAP values quantify the direction and magnitude of the impact of input features on model predictions, with the horizontal axis representing their influence on prediction outcomes in terms of strength and direction. The color gradient reflects the magnitude of feature values—blue indicates low values and red represents high values, while each dot corresponds to the feature contribution of an individual sample within the dataset. Figure 5a, based on SHAP analysis of OS features, reveals that EEVI, CRSI, and Elevation exhibit high feature importance. Notably, the two most influential features (EEVI, CRSI) are vegetation indices, indicating that composite spectral indices outperform individual spectral bands. High EEVI values (red points) are concentrated within the negative SHAP region. For both EEVI and CRSI, high feature values predominantly impact lower model predictions, whereas low feature values mainly influence higher model predictions. Importantly, the magnitude of influence exerted by low values is greater than that exerted by most high values. SHAP analysis incorporating time-series vegetation features reveals that NDVI_max, NDVI_mean, and Elevation exhibit high importance, with NDVI_max reaching an importance value of 0.3. The two highest-ranking features (NDVI_max and NDVI_mean) are both time-series-integrated vegetation indices, while the importance of single-temporal indices EEVI and ERSSI decreases drastically (Figure 5b). For NDVI_max, most high feature values exert influence toward lower model outputs, whereas low feature values primarily impact higher model outputs. This pattern aligns with the physiological mechanism whereby salinity accumulation suppresses peak vegetation biomass. Moreover, the magnitude of impact exerted by low values exceeds that produced by most high NDVI_max values. SHAP analysis reveals a robust negative correlation for NDVI_max, evidenced by the concentration of its high values within negative SHAP regions and the wider dispersion of its low values across positive SHAP regions. Elevation exhibits a complex, non-monotonic relationship. Crucially, Elevation consistently emerges as a decisive factor across both modeling strategies (spectral-feature-only and combined spectral-temporal features), confirming the fundamental role of hydro-topographic parameters in governing salinity accumulation processes regardless of the input feature configuration. Furthermore, the incorporation of topographic factors substantially enhanced the model’s explanatory power.

3.5. Soil Salinity Maps

Figure 7 systematically compares the spatial distribution patterns of soil salinity derived from three machine learning models (CatBoost, XGBoost, and LightGBM) under two feature strategies—OS (Optical Spectral features only) and OVTS (integrated Optical-Vegetation Time-Series features)—using monthly composite Sentinel-2 imagery from 2024 to spring 2025. In the OS-based models (a–c), scattered high-salinity areas are observed, primarily concentrated in the northern, northwestern, and central coastal regions, with highly saline soils predominantly accumulating in littoral and nearshore zones. These models are characterized by a wide predicted salinity range and a noticeable tendency toward systematic overestimation.In contrast, the OVTS-based models (d–f) exhibit finer spatial details and more reasonable salinity gradients. High-salinity zones are rendered as more fragmented and reduced in spatial extent, which significantly mitigates the overestimation observed in the OS models. The OVTS strategy is shown to consistently improve spatial continuity and physical plausibility across all models. In particular, the LightGBM-OVTS combination (Figure 7f) is demonstrated to effectively suppress abnormal high-value patches while preserving spatial detail.These results confirm that the incorporation of vegetation time-series features (OVTS) enhances the representation of soil salinity spatial patterns in complex coastal environments by capturing vegetation response mechanisms to salt stress, thereby establishing a reliable methodological foundation for precision monitoring of regional soil salinization.

4. Discussion

4.1. Influence of Vegetation Information on Soil Salinity Mapping

Soil salinity exhibits seasonal variation patterns [64]. In the Yellow River Delta, vegetation types and their distribution are significantly correlated with soil salinity. Research by Wang et al. (2017) indicates that vegetation tolerant to high salinity includes seepweed (Suaeda salsa), common reed (Phragmites australis), cogon grass (Imperata cylindrica), and cotton. In contrast, low salinity areas are primarily cultivated with rice and winter wheat [65]. The dominant cropping system in this region, characterized by wheat-maize rotation, follows a sequential pattern of farming activities such as tillage and irrigation.

Figure 8 presents the correlation coefficients between the Normalized Difference Vegetation Index (NDVI) at different time phases from 2024 to 2025 and spring soil salinity content. A lagged response mechanism is observed, with soil salinity showing the strongest correlation with the maximum NDVI (NDVI_max) and average NDVI (NDVI_mean) from July to September of the previous year, which is primarily attributed to crop phenology and regional farming practices [37,50]. By extracting key statistical metrics (minimum, maximum, mean) of NDVI within different time windows to characterize its temporal features, the role of vegetation dynamics in soil salinity inversion is revealed in this study.Among the 119 soil sampling points, 72 were located in areas covered by winter wheat, crop residues, and deep-tilled soil, while the remaining 47 were situated in saline-alkali bare land and sparse vegetation. As shown in Figure 9, low NDVI values (<0.2) during the sampling period are found to coexist with low salinity levels. The underlying mechanism is explained by physical disturbance: as illustrated in Figure 2e–f, deep tillage incorporates crop residues (e.g., straw) into the topsoil (0–20 cm). Crop residue coverage is shown to reduce soil reflectivity, while mechanical disturbances such as deep tilling alter surface roughness, resulting in abnormal backscattering. Consequently, the correlation between concurrent NDVI and topsoil salinity is weakened.

Tillage activities are identified as key interference sources affecting spring salinity inversion, whereas the temporal characteristics of NDVI are demonstrated to exhibit a stronger correlation with salinity [66]. Agricultural activities disrupt the physical linkage between NDVI and salinity in traditional single-phase methods, leading to overestimation of low-salinity areas by the OS model in Figure 5a–c and reduced accuracy in concurrent soil salinity inversion. As emphasized by Whitney [67], reliance solely on imagery synchronized with sampling time or single-phase remote sensing is insufficient for soil salinization inversion. Single-phase methods are unable to capture the cumulative effects of seasonal dynamic interactions within the soil-vegetation-water system [37]. These findings are consistent with the academic consensus that exclusive dependence on field-sampling synchronization or single-phase remote sensing imagery cannot effectively invert salinity levels [22,68]. Therefore, the temporal characteristics of NDVI are established as critical indicators for characterizing soil desalination status and resistance to salt accumulation. The time-series features of NDVI integrate the cumulative effects of crop rotation systems, irrigation patterns, and land management strategies—elements essential for maintaining salt balance in the soil profile—making them strong predictors of spring salinization levels. A significant negative correlation is revealed in Figure 8 between the previous summer’s (July–September) NDVI_max and spring salinity. Mechanistic studies indicate that vigorous transpiration during the peak growth stage of summer maize promotes root water uptake, thereby lowering groundwater levels and inhibiting the capillary rise in spring salinity. Simultaneously, summer precipitation combined with artificial irrigation enhances salt leaching in the root zone [69], effectively reducing surface salt accumulation caused by soil evaporation. This compound effect significantly reduces topsoil salinity, with its impact often lasting until the following spring under continuous vegetation cover and suitable hydrothermal conditions, as leached salts are less prone to rapid reaccumulation [68]. The vegetation phenological features extracted by OVTS (e.g., NDVI_max) not only avoid spring tillage interference but also capture the inhibitory effect on salt accumulation, significantly enhancing model robustness (Figure 5d–f).

Ultimately, predictive variables for soil salinity mapping were successfully screened from spectral indices and vegetation temporal features through the combined application of Pearson correlation analysis and the Variable Importance in Projection (VIP) algorithm. Consequently, the OVTS features derived from Sentinel-2 imagery are demonstrated to provide supplementary information for salinity inversion, capturing both the hysteresis effect of salinity and effectively avoiding interference from spring tillage. These features are found to contain unique salinity information that is inaccessible to other predictors, significantly enhancing model accuracy. In this study, the accuracy achieved by the LightGBM model (R² = 0.77) is attributed to the multi-temporal phenological features that were introduced, which is consistent with the previous findings [40], indicating that the proposed framework has a certain degree of universality for soil salinity inversion in irrigation districts. However, the direct incorporation of the complete NDVI time series into the model is considered impractical due to autocorrelation and redundancy between adjacent NDVI values. To reduce temporal dimensionality redundancy, Principal Component Analysis (PCA) was employed and aggregated temporal statistics (maximum, minimum, mean) were calculated. By capturing phenological response processes and analyzing the salinity regulation effects of the wheat-maize rotation system, the limitations of single-phase imagery are overcome by the OVTS features. The association between vegetation temporal characteristics and soil salinity is significantly strengthened, ensuring robust inversion performance even in spatiotemporal domains dominated by non-photosynthetic vegetation.

4.2. Research Uncertainties and Limitations

This study is also subject to several limitations. First, extensive vegetation cover is observed across most of the Yellow River Delta, resulting in minimal exposed bare soil during the study period. At the time of sampling, which coincided with the green-up phase of winter crops, the acquisition of pure soil spectra was largely unattainable at most sampling sites. Furthermore, multiple factors are identified as potential exacerbators of uncertainty in the relationship between remotely sensed data and soil salinity. Beyond soil salinity, vegetation characteristics are also influenced by climate variations, anthropogenic activities, and other soil properties, introducing confounding effects on spectral signals. In some agricultural fields, crop residues are incorporated into the surface soil through spring tillage practices. These surface covers are found to significantly diminish the sensitivity of salinity indices to underlying soil salt content, particularly during the critical period immediately preceding the completion of spring tillage.

Although surface soil salinity (0–20 cm depth) was determined via conventional methods at 119 sampling sites within the Yellow River irrigation districts, the salinity status of deeper soil layers is also considered to impact vegetation growth, representing an unquantified variable. The region is simultaneously affected by dual influences of seawater intrusion and agricultural activities, such as spring tillage and irrigation. During model application, the practical difficulty of completely excluding pixels containing mountain shadows, water bodies, and built-up areas is recognized to introduce further uncertainty into the retrieval results.

Constrained by data availability, key hydrological elements—including groundwater depth and irrigation water usage—were not incorporated into the model. The integration of such data is expected to further enhance model accuracy. In future work, groundwater monitoring data will be coupled with irrigation records to mechanistically quantify the impact of agricultural activities on the accuracy of salinity inversion.

5. Conclusions

This study confirms that utilizing multi-temporal Sentinel satellite data and integrating multi-temporal vegetation phenological features can significantly enhance the accuracy of soil salinity mapping in coastal irrigation districts. The core conclusions are as follows:

• The proposed OVTS framework derived from Savitzky-Golay-filtered NDVI time-series (PCs and extremum features) effectively captures vegetation response to salt stress, overcoming spectral interference from agricultural activities (e.g., spring tillage residue coverage). This reduces low-salinity overestimation compared to single-temporal approaches.
• The LightGBM model integrating spectral and vegetation temporal features within the OVTS framework significantly enhanced the spatial coherence of salinity gradients and effectively improved spatial refinement capabilities. It achieved optimal validation accuracy (R² = 0.77, RMSE = 0.26 g/kg), with a 13% increase in R² and a 27.7% reduction in RMSE compared to spectral only models.
• SHAP analysis reveals that vegetation factors and topographic factors serve as key predictors in both framework strategies. When vegetation temporal features are incorporated, they emerge as crucial predictive factors that effectively mitigate strong interference from spring tillage and residue cover, thereby maximizing the signal-to-noise ratio. The July–September vegetation characteristic window was identified as the optimal period for remote sensing-based inversion of soil salinity. The peak vegetation biomass during this period demonstrates a significant negative correlation with soil salinity levels in the following spring, indicating a clear lagged inhibitory effect of vegetation physiological activity on salt accumulation during this timeframe.