Research on Soil Salinity Inversion in Coastal Areas Based on UAV Multispectral Imagery and Ensemble Machine Learning

Abstract

Accurate and timely monitoring of soil salinity is of great significance for the ecological restoration of saline-alkali land and precision agricultural management. In this study, a typical coastal saline-alkali farmland located in Huanghua City, Hebei Province, China, in the Bohai coastal region, was selected as the study area. High-resolution images were acquired using an unmanned aerial vehicle (UAV) equipped with a multispectral sensor, and ground soil salinity samples were collected synchronously. Based on the construction of a feature library comprising spectral reflectance, vegetation indices, and salinity indices, three algorithms, PSO-SFLA, MultiSURF, and VIP, were employed for feature selection. Subsequently, an ensemble model was established, utilizing Ridge Regression (Ridge), Random Forest (RF), and Extra Trees (ET) as primary base learners, and Extreme Gradient Boosting (XGBoost) as the secondary meta-learner. This ensemble model was applied for soil salinity inversion. Furthermore, the coefficient of determination (R²), standardized root mean square error (S_RMSE), and the ratio of performance to interquartile distance (RPIQ) were introduced to comprehensively evaluate the accuracy of the models. Finally, the intrinsic physical responses of the features were explored through SHAP. The results showed that the optimization by the PSO-SFLA effectively reduced the impact of spectral multicollinearity, and 11 core features highly sensitive to salinity were selected from a vast number of indices. The ensemble model showed better predictive performance on the independent test set, achieving an R² of 0.758, an S_RMSE of 0.285, and an RPIQ of 3.382, outperforming the single Ridge, RF, and ET models under the current experimental conditions. Based on this model, the spatial distribution map of soil salinity in the experimental area was generated. The integrated and interpretable workflow proposed in this study, combining UAV multispectral imagery, PSO-SFLA-based feature selection, ensemble learning, and SHAP interpretation, provides a practical approach for accurate soil salinity inversion and dynamic agricultural monitoring in coastal saline-alkali lands.

1. Introduction

Soil salinization is a widespread and increasingly severe form of land degradation globally, which severely restricts the health of terrestrial ecosystems and the sustainable development of agriculture [1]. Approximately 33% of the world’s land surface has been degraded [2]. As a prevalent form of soil degradation globally, soil salinization affects 466.36 million ha of surface soil and 377.68 million ha of subsoil in Asia alone [3,4]. The area of salt-affected soils covers approximately 100 million hectares in China, with over 80% remaining undeveloped [5]. It continuously threatens crop growth and species diversity; therefore, the real-time dynamic monitoring of soil salinization is of paramount importance for food security and the development of ecological agriculture [6]. Traditional soil salinity monitoring heavily relies on manual sampling using field soil augers and subsequent laboratory physicochemical analyses. Although this observation method provides extremely high calibration accuracy, it is time-consuming, labor-intensive, and costly [7,8]. Furthermore, the high spatial heterogeneity of soil salinity across large-scale farmlands is difficult to characterize by this approach. With the rapid development of remote sensing technology, satellite remote sensing, represented by Landsat and Sentinel, has been widely applied due to its broad swath coverage [9,10]. However, constrained by the mixed-pixel problem caused by low spatial resolution, and further restricted by susceptibility to cloud and rain obstruction as well as long revisit cycles, satellite remote sensing struggles to satisfy the demands of modern precision agriculture for continuous and micro-scale surface observation [11]. In recent years, low-altitude unmanned aerial vehicle (UAV) remote sensing platforms have compensated for these technical limitations by virtue of their flexibility, ultra-high spatial resolution, and low operational costs. Consequently, UAV platforms have become an effective tool for monitoring the micro-scale spatial and temporal patterns of soil salinity in farmlands [12].

Currently, methods for soil salinity inversion based on spectral information are primarily categorized into empirical approaches and machine learning techniques [13]. Early or fundamental empirical methods typically involve directly selecting the red and near-infrared bands that are highly sensitive to salinity, or constructing vegetation indices and soil salinity indices through algebraic operations to enhance specific spectral absorption features. Subsequently, empirical equations are established by incorporating statistical techniques such as linear regression or multiple stepwise regression [14]. However, the relationships among micro-scale soil salinity, moisture, and surface reflectance are extremely complex. Empirical methods mainly rely on linear relationships, which oversimplify the actual physical processes, and it is inadequate for comprehensively simulating the nonlinear dynamics between salinity and spectral properties, which often leads to limited monitoring accuracy [15]. Consequently, machine learning methods, equipped with powerful nonlinear fitting and high-dimensional data processing capabilities, have progressively replaced traditional regression approaches, emerging as the mainstream paradigm for extracting soil salinization [16]. Advanced machine learning algorithms can handle complex datasets and capture nonlinear relationships. For instance, Random Forest [17] (RF) can flexibly model complex and deep internal correlations through its tree ensemble structure, while Support Vector Machine [18] (SVM) and similar algorithms exhibit outstanding performance in the regression and classification of multispectral nonlinear problems based on kernel function hyperplane mapping mechanisms. Compared with traditional empirical methods, these machine learning algorithms demonstrate superior robustness and generalization abilities, and they have achieved significant success in inversion applications within arid and complex coastal saline-alkali environments [13,14,15,16,17,18,19].

Although machine learning exhibits significant advantages in handling multivariate data, the input feature set is inevitably expanded to enhance weak salinity signals, typically by stacking a massive number of vegetation indices, salinity indices, or multi-source channel features. This practice is highly prone to triggering the curse of dimensionality, introducing severe information redundancy and multicollinearity issues [20]. Excessive redundant variables not only increase the computational cost of models but also obscure the true contributions of specific bands, thereby degrading the efficiency of model construction and the ultimate prediction accuracy. Therefore, the scientific and prudent selection of the number and types of input variables is of critical importance. Traditional feature selection methods (e.g., correlation analysis, principal component analysis, variable importance in projection, or elastic net algorithms) are mostly constrained to capturing linear features or achieving superficial dimensionality reduction. When confronted with the deep, nonlinear, interactive correlations between soil salinity and multispectral signals, their feature selection capabilities are often highly limited.

Furthermore, the process of salt accumulation and leaching in saline soils exhibits exceedingly strong spatiotemporal variability [21]. Influenced by this intense spatial variability, limited predictive capabilities are often exhibited by single machine learning methods when confronting complex saline-alkali patches across varying surface environmental gradients. Such single models are highly susceptible to prediction distortion [22,23], and it is difficult for them to maintain high and stable monitoring accuracy at regional or field scales. Therefore, the stability and predictive capability of models can be enhanced to a certain extent by combining the advantages of multiple independent machine learning algorithms. Ensemble learning is a widely adopted collective learning technique that has been extensively applied across various machine learning tasks. Superior and more comprehensive supervised learning performance can be achieved through the integration of multiple base learners. In agricultural research, complex surface mixed interferences were successfully isolated by Das et al. [24] using advanced spectral techniques combined with ensemble machine learning, which significantly improved the monitoring accuracy of soil salinity content. Moreover, it was demonstrated by Wang et al. [25] that when utilizing high-resolution UAV imagery to estimate heterogeneous soil salinity, the robustness and coefficient of determination of a feature-optimized weighted ensemble learning model far exceeded those of traditional single baseline models. However, current research on constructing such multi-level ensemble architectures in the field of coastal soil salinity inversion remains relatively limited. Simultaneously, the “black-box” nature of models is severely exacerbated by highly integrated model architectures, which hinders their practical wide application and deep mechanistic interpretation. To overcome this challenge, the Shapley Additive Explanations (SHAP) method [26], rooted in game theory, has been gradually introduced into the interpretation of agricultural remote sensing models in recent years. The absolute marginal impacts of input feature variables on the model output can be comprehensively quantified by this method from both global and local perspectives, thereby providing pure data-driven models with a solid physical and agronomic foundation. Specifically, the black-box decision-making process of a coastal farmland salinity prediction model was successfully elucidated by Jia et al. [13] using the SHAP interpretability framework. This application not only quantitatively verified the nonlinear physiological response mechanisms between specific spectral index feature thresholds and severe soil salinity stress, but also ensured the absolute internal logical self-consistency of the ensemble algorithm.

In summary, UAV-based soil salinity inversion in coastal saline-alkali areas still faces three key challenges. First, the strong spatial heterogeneity of soil salinity makes it difficult for a single prediction model to maintain stable performance under complex field conditions. Second, the construction of numerous vegetation and salinity indices from UAV multispectral imagery inevitably introduces redundancy and multicollinearity among input features. Third, although ensemble machine-learning models can improve prediction accuracy, their complex structures also increase the difficulty of interpreting the physical meaning of model decisions. To address these challenges, this study develops an integrated and interpretable workflow for UAV-based coastal soil salinity inversion. The workflow combines PSO-SFLA-based feature selection, stacking ensemble learning, and SHAP-based interpretation. The main contributions of this study are as follows: (1) the performance of VIP, MultiSURF, and PSO-SFLA feature selection strategies is systematically compared under the same UAV multispectral dataset; (2) the PSO-SFLA-selected compact feature subset is used to reduce feature redundancy and improve prediction reliability; and (3) SHAP analysis is introduced to explain the contributions of both base learners and selected spectral features. This workflow provides a practical reference for high-resolution soil salinity mapping and dynamic monitoring in coastal agricultural areas.

2. Materials and Methods

2.1. Study Area

Huanghua City, located in Hebei Province, northern China, is situated in the Bohai coastal plain, with geographical coordinates ranging from 117°05′ E to 117°49′ E and 38°09′ N to 38°39′ N. It is affiliated with the Heilonggang River Basin. The municipality covers a total area of 2212 km², within which an arable land area of 613.07 km² and a total saline-alkali land area of 235.93 km² are encompassed. Situated in a warm-temperate semi-humid continental monsoon climate zone, slight maritime climate characteristics are exhibited by the city due to its proximity to the Bohai Sea. It is characterized by significant monsoons and four distinct seasons. A multi-year average sunshine duration of 2700 h, an annual precipitation of approximately 627 mm, while the mean annual evaporation is approximately 1800 mm. The groundwater depth ranges from 1.0 to 1.5 m, with a mineralization degree greater than 3 g/L. The tested soil is a coastal chloride saline soil with a silty loam texture. Furthermore, approximately 65% of the precipitation is concentrated in July and August. Influenced by marine depositional landforms, a high salinity of 10 to 30 g/L is found in shallow groundwater. Additionally, deep freshwater must be extracted from depths of 250 to 800 m; consequently, immense difficulties and high costs are associated with its development and utilization. A relatively monotonous crop composition and a stable planting structure are maintained in the study area, where a crop rotation system of winter wheat and summer maize is predominantly implemented year-round. The location of the experimental site is shown in Figure 1.

2.2. Field Data Collection and Analysis

Field sampling was conducted in winter wheat fields located in Lizizha Village, Changguo Township, Huanghua City, on 4 December 2025. No snow cover was present on the soil surface during data collection, ensuring that the multispectral reflectance information represented the actual surface soil conditions. To improve the spatial representativeness of soil sampling and reduce excessive local clustering of sampling points, a grid-based random sampling strategy was designed using ArcGIS. The sampling area was first divided into regular grid cells, and 90 sampling sites were then distributed within the grid framework to ensure that the samples covered the main spatial extent of the study area. This sampling design helped obtain a spatially balanced dataset and reduced the potential bias caused by the over-concentration of samples in limited local areas. Soil samples at a depth of 0–10 cm were collected from these sites using a soil auger, resulting in a total of 90 soil samples, and the geographic coordinates of each sampling site were recorded using a handheld GPS device (ZL Electronic Technology, Anhui, China). The collected soil samples were dried in an oven at 105 °C for 8 h, after which their dry weights were recorded. Subsequently, each dried sample was ground and sieved; a 20 g subsample was then mixed with 100 mL of distilled water to prepare a 1:5 soil-water suspension. The solution was thoroughly stirred with a glass rod and allowed to stand to ensure complete reaction. Ultimately, the supernatant was filtered, and its electrical conductivity (EC_1:5) was measured utilizing a DDS-307A conductivity meter (Shanghai INESA Scientific Instrument CO., LTD, Shanghai, China), with the unit expressed as μS/cm.

2.3. Acquisition of UAV Multispectral Remote Sensing Imagery

The remote sensing platform utilized in the experiment was the Phantom 4 Multispectral UAV, manufactured by DJI Innovation Technology Co., Ltd (Shenzhen, China). It was synchronously equipped with six 1/2.9-inch CMOS image sensors, comprising one color RGB camera channel and five monochrome channels: blue (450 nm), green (560 nm), red (650 nm), red edge (730 nm), and near-infrared (840 nm), as shown in Figure 2. The acquisition time of the UAV multispectral imagery was synchronized with the ground soil sampling on 4 December 2025. The UAV flight was conducted at approximately 12:00 noon, when solar illumination was relatively stable, and shadow effects were minimized. During the experiment, clear weather conditions with minimal wind were observed. Since the field campaign was carried out during the winter wheat seedling stage, the projective vegetation cover was relatively low, and most of the soil surface remained exposed. A flight altitude of 80 m was maintained by the UAV, with the multispectral camera lens oriented vertically downward. Both the forward overlap and side overlap rates were set to 80% and 70%, respectively. Prior to the flight, image calibration was executed using a standard whiteboard. The acquired multispectral images were imported into Pix4Dmapper 4.5.6, where geometric correction, radiometric calibration, and mosaicking processing were performed to obtain a complete orthophoto map. The final UAV multispectral orthomosaic had an average ground sampling distance (GSD) of 2.48 cm/pixel (0.98 in/pixel), providing high-resolution spatial information for extracting reflectance values at the soil sampling locations. By utilizing ArcGIS 10.8 software, the latitude and longitude coordinates of the sampling points were imported, thereby extracting the corresponding reflectance values.

2.4. Construction and Selection of Spectral Indices

Information regarding geomorphic features can be enhanced by combinations of different spectral bands. Spectral indices are indicators that comprehensively consider the spectral characteristics of various bands of ground objects, applying mathematical transformations and combinations to the reflectance of these bands to enhance specific informational features. According to recent remote sensing observational studies, it was discovered that the red edge band represents a sensitive transition zone between strong absorption by plant canopy chlorophyll and strong reflection in the near-infrared region. It is extremely sensitive to weak physiological fluctuations induced by salinity stress in the underlying soil [27]. Therefore, by applying an allelic substitution of the red band—which is susceptible to atmospheric and soil background interference in traditional vegetation or salinity indices—with the red edge band, improved spectral indices with higher robustness against multi-source severe salinity stress can be derived [28]. In this study, 5 original bands, 26 traditional spectral indices, and 6 improved spectral indices were selected for screening; the calculation formulas are listed in

Table 1. Calculation formula of spectral index.

Spectral Index	Computational Formula	Reference
Normalized salt index (NDSI)	$\mathrm{N}\mathrm{D}\mathrm{S}\mathrm{I}=(\mathrm{R}-\mathrm{NIR})/(\mathrm{R}+\mathrm{N}\mathrm{I}\mathrm{R})$	[29]
Improved normalized salt index (NDSI-reg)	$\mathrm{N}\mathrm{D}\mathrm{S}\mathrm{I}-\mathrm{r}\mathrm{e}\mathrm{g}=(\mathrm{R}\mathrm{e}\mathrm{d}\mathrm{E}\mathrm{d}\mathrm{g}\mathrm{e}-\mathrm{N}\mathrm{I}\mathrm{R})/(\mathrm{R}\mathrm{e}\mathrm{d}\mathrm{E}\mathrm{d}\mathrm{g}\mathrm{e}+\mathrm{N}\mathrm{I}\mathrm{R})$	[30]
Salinity Index 1 (S1)	$\mathrm{S}1=\mathrm{B}/\mathrm{R}$	[29]
Salinity Index 2 (S2)	$\mathrm{S}2=(\mathrm{B}-\mathrm{R})/(\mathrm{B}+\mathrm{R})$	[29]
Salinity Index 3 (S3)	$\mathrm{S}3=(\mathrm{G}\times \mathrm{R})/\mathrm{B}$	[29]
Salinity Index 4 (S4)	$\mathrm{S}4=\sqrt{\mathrm{B}\times \mathrm{R}}$	[29]
Salinity Index 5 (S5)	$\mathrm{S}5=(\mathrm{B}\times \mathrm{R})/\mathrm{G}$	[29]
Salinity Index 6 (S6)	$\mathrm{S}6=(\mathrm{R}\times \mathrm{N}\mathrm{I}\mathrm{R})/\mathrm{G}$	[29]
Salinity Index 1 (SI1)	$\mathrm{S}\mathrm{I}1=\sqrt{\mathrm{G}\times \mathrm{R}}$	[29]
Salinity Index 2 (SI2)	$\mathrm{S}\mathrm{I}2=\sqrt{{\mathrm{G}}^2+{\mathrm{R}}^2+{\mathrm{N}\mathrm{I}\mathrm{R}}^2}$	[29]
Salinity Index 3 (SI3)	$\mathrm{S}\mathrm{I}3=\sqrt{{\mathrm{G}}^2+{\mathrm{R}}^2}$	[29]
Salinity Index-T (SI-T)	$\mathrm{S}\mathrm{I}-\mathrm{T}=100(\mathrm{R}-\mathrm{N}\mathrm{I}\mathrm{R})$	[29]
Improved salinity index 1 (SI1-reg)	$\mathrm{S}\mathrm{I}1-\mathrm{r}\mathrm{e}\mathrm{g}=\sqrt{\mathrm{G}\times \mathrm{R}\mathrm{e}\mathrm{d}\mathrm{E}\mathrm{d}\mathrm{g}\mathrm{e}}$	[30]
Improved salinity index 2 (SI2-reg)	$\mathrm{S}\mathrm{I}2-\mathrm{r}\mathrm{e}\mathrm{g}=\sqrt{{\mathrm{G}}^2+{\mathrm{R}\mathrm{e}\mathrm{d}\mathrm{E}\mathrm{d}\mathrm{g}\mathrm{e}}^2+{\mathrm{N}\mathrm{I}\mathrm{R}}^2}$	[30]
Improved salinity index 3 (SI3-reg)	$\mathrm{S}\mathrm{I}3-\mathrm{r}\mathrm{e}\mathrm{g}=\sqrt{{\mathrm{G}}^2+{\mathrm{R}\mathrm{e}\mathrm{d}\mathrm{E}\mathrm{d}\mathrm{g}\mathrm{e}}^2}$	[30]
Brightness Index (BI)	$\mathrm{B}\mathrm{I}=\sqrt{{\mathrm{R}}^2+{\mathrm{N}\mathrm{I}\mathrm{R}}^2}$	[29]
Normalized Difference Vegetation Index (NDVI)	$\mathrm{N}\mathrm{D}\mathrm{V}\mathrm{I}=(\mathrm{N}\mathrm{I}\mathrm{R}-\mathrm{R})/(\mathrm{N}\mathrm{I}\mathrm{R}+\mathrm{R})$	[31]
Improved Normalized Difference Vegetation Index (NDVI-reg)	$\mathrm{N}\mathrm{D}\mathrm{V}\mathrm{I}-\mathrm{r}\mathrm{e}\mathrm{g}=(\mathrm{N}\mathrm{I}\mathrm{R}-\mathrm{R}\mathrm{e}\mathrm{d}\mathrm{E}\mathrm{d}\mathrm{g}\mathrm{e})/(\mathrm{N}\mathrm{I}\mathrm{R}+\mathrm{R}\mathrm{e}\mathrm{d}\mathrm{E}\mathrm{d}\mathrm{g}\mathrm{e})$	[30]
Difference Vegetation Index (DVI)	$\mathrm{D}\mathrm{V}\mathrm{I}=\mathrm{N}\mathrm{I}\mathrm{R}-\mathrm{R}$	[31]
Rededge Difference Vegetation Index (DVI-reg)	$\mathrm{D}\mathrm{V}\mathrm{I}-\mathrm{r}\mathrm{e}\mathrm{g}=\mathrm{N}\mathrm{I}\mathrm{R}-\mathrm{R}\mathrm{e}\mathrm{d}\mathrm{E}\mathrm{d}\mathrm{g}\mathrm{e}$	[30]
Enhanced Vegetation Index (EVI)	$\mathrm{E}\mathrm{V}\mathrm{I}=2.5(\mathrm{N}\mathrm{I}\mathrm{R}-\mathrm{R})/(\mathrm{N}\mathrm{I}\mathrm{R}+6\mathrm{R}-7.5\mathrm{B}+1)$	[32]
Triangular Vegetation Index (TVI)	$\mathrm{T}\mathrm{V}\mathrm{I}=60\left(\mathrm{N}\mathrm{I}\mathrm{R}-\mathrm{G}\right)-100(\mathrm{R}-\mathrm{G})$	[31]
Soil Adjusted Vegetation Index (SAVI)	$\mathrm{S}\mathrm{A}\mathrm{V}\mathrm{I}=1.5(\mathrm{N}\mathrm{I}\mathrm{R}-\mathrm{R})/(\mathrm{N}\mathrm{I}\mathrm{R}+\mathrm{R}+0.5)$	[32]
Normalized Difference Greenness Vegetation Index (NDGI)	$\mathrm{N}\mathrm{D}\mathrm{G}\mathrm{I}=(\mathrm{G}-\mathrm{R})/(\mathrm{G}+\mathrm{R})$	[33]
Ratio Vegetation Index (RVI)	$\mathrm{R}\mathrm{V}\mathrm{I}=\mathrm{N}\mathrm{I}\mathrm{R}/\mathrm{R}$	[31]
Optimized Soil Adjusted Vegetation Index (OSAVI)	$\mathrm{O}\mathrm{S}\mathrm{A}\mathrm{V}\mathrm{I}=1.16(\mathrm{N}\mathrm{I}\mathrm{R}-\mathrm{R})/(\mathrm{N}\mathrm{I}\mathrm{R}+\mathrm{R}+0.16)$	[32]
Modified Chlorophyll Absorption Reflectance Index (MCARI)	$\mathrm{M}\mathrm{C}\mathrm{A}\mathrm{R}\mathrm{I}=\left[\mathrm{R}\mathrm{e}\mathrm{d}\mathrm{E}\mathrm{d}\mathrm{g}\mathrm{e}-\mathrm{R}-0.2\left(\mathrm{R}\mathrm{e}\mathrm{d}\mathrm{E}\mathrm{d}\mathrm{g}\mathrm{e}-\mathrm{G}\right)\right]\times \mathrm{R}\mathrm{e}\mathrm{d}\mathrm{E}\mathrm{d}\mathrm{g}\mathrm{e}/\mathrm{R}$	[34]
Green Normalized Difference Vegetation Index (GNDVI)	$\mathrm{G}\mathrm{N}\mathrm{D}\mathrm{V}\mathrm{I}=(\mathrm{N}\mathrm{I}\mathrm{R}-\mathrm{G})/(\mathrm{N}\mathrm{I}\mathrm{R}+\mathrm{G})$	[34]
Modified Soil-adjusted Vegetation Index (MSAVI)	$\mathrm{M}\mathrm{S}\mathrm{A}\mathrm{V}\mathrm{I}=\left[\left(2\mathrm{N}\mathrm{I}\mathrm{R}-1\right)-\sqrt{{\left(2\mathrm{N}\mathrm{I}\mathrm{R}+1\right)}^2-(8\mathrm{N}\mathrm{I}\mathrm{R}-\mathrm{R})}\right]/2$	[32]
Re-normalized Difference Vegetation Index (RDVI)	$\mathrm{R}\mathrm{D}\mathrm{V}\mathrm{I}=(\mathrm{N}\mathrm{I}\mathrm{R}-\mathrm{R})/\sqrt{\mathrm{N}\mathrm{I}\mathrm{R}+\mathrm{R}}$	[32]
Weight Difference Vegetation Index (WDVI)	$\mathrm{W}\mathrm{D}\mathrm{V}\mathrm{I}=\mathrm{N}\mathrm{I}\mathrm{R}-1.06\mathrm{R}$	[32]
Infrared Percentage Vegetation Index (IPVI)	$\mathrm{I}\mathrm{P}\mathrm{V}\mathrm{I}=\mathrm{N}\mathrm{I}\mathrm{R}/(\mathrm{N}\mathrm{I}\mathrm{R}+\mathrm{R})$	[32]

Note: R, G, B, NIR, and RedEdge represent the reflectance of the red, green, blue, near-infrared, and red edge bands, respectively.

.

2.5. Data Analysis and Modeling Methods

2.5.1. Data Partitioning and Validation Strategy

To improve the reproducibility of the modeling workflow and reduce the risk of overfitting and information leakage, data preprocessing, partitioning, and validation strategies were determined before feature selection and model construction. After the construction of the initial 37 candidate variables, invalid variables with zero variance were removed. Missing values and non-finite values generated during spectral index calculation were also checked and removed before modeling. In addition, ground-measured EC1:5 samples with potential extreme values were examined using the ±3 standard deviation criterion. Samples identified as abnormal measurement errors were excluded to reduce disproportionate effects on model training.

The soil samples were then divided into a training set and an independent testing set at a ratio of 7:3. To ensure that both subsets represented the full range of soil salinity variation, stratified random sampling based on target EC quantiles was adopted. The training set was used for feature selection, parameter optimization, cross-validation, and model training, whereas the independent testing set was kept separate and used only for final model evaluation. All feature selection procedures were performed within the training set to avoid information leakage from the testing set.

Because the original reflectance bands and spectral indices had different numerical ranges, Z-score standardization was applied to the input feature matrix where required by the model. To avoid information leakage, the scaling parameters were fitted only on the training set and were then applied to the independent testing set. The same preprocessing procedure was used across all feature selection methods and prediction models to ensure a fair comparison.

Five-fold cross-validation was applied within the training set during model development. The same training/testing partition and cross-validation folds were used for all feature selection methods and prediction models. To improve the reproducibility of stochastic procedures, a fixed random seed was used throughout the workflow. The random seed was set to 42 for data partitioning, cross-validation splitting, feature selection, and model training procedures involving stochastic operations, including RF, ET, XGBoost, and PSO-SFLA.

2.5.2. Feature Selection

The spectral features extracted from multispectral images often contain high dimensionality, redundancy, and multicollinearity. To effectively eliminate the noise impact from the environmental background and remove redundant bands exhibiting multicollinearity, three feature selection methods were introduced and compared in this study to select the optimal variables from the spectral data:

Variable Importance in Projection (VIP) technique based on linear variance projection [33]: This method is rooted in Partial Least Squares (PLS) regression and primarily assigns a weighted shared importance to features by computing the degree to which each independent variable explains the variance in a multi-dependent variable model. The prevalent standard was adhered to in this study, whereby variables with VIP scores greater than 1 were extracted as the cutoff filtering indicator to eliminate superficial and useless band features.

MultiSURF algorithm based on permutation test hypothesis [35]: As an advanced evolutionary extension of the Relief algorithmic family, MultiSURF evaluates features based on their ability to distinguish nearest neighbor samples of the same class from those of different classes. Its core procedure adopts a robust permutation testing strategy to construct a probabilistic null hypothesis distribution model for feature selection. A target feature subset is then extracted at a significance level threshold of α = 0.05 to ensure the statistical significance of the selected features.

PSO-SFLA deep dimensionality reduction wrapper optimization based on collaborative Particle Swarm Optimization and Shuffled Frog Leaping [36]: This module integrates not only the velocity-position update formula characteristics of Particle Swarm Optimization (PSO) for sensitive drifting in large solution spaces, but also blends the advantages of deep, multi-population competition and local cultural exchange, which the Shuffled Frog Leaping Algorithm (SFLA) excels at [37]. In terms of operational configuration, XGBoost was utilized as the baseline evaluator at the fundamental level of the PSO-SFLA, and the standardized root mean square error (S_RMSE) convergence value returned by 5-fold cross-validation was employed as the fitness function for the search. Furthermore, this module was additionally nested within a stability validation framework; through multiple independent, parallel executions, the appearance frequencies of various spectral variables were scored and summarized. Ultimately, the features selected most frequently were retained as the optimal subset. It should be noted that XGBoost was used only as a nonlinear fitness evaluator during the wrapper-based PSO-SFLA search process, rather than as the only final prediction model. The purpose of using XGBoost was to provide a robust nonlinear evaluation of candidate feature subsets under cross-validation. To reduce potential model-structure bias, the selected feature subset was subsequently evaluated using multiple independent inversion models, including Ridge, RF, ET, XGBoost, and the ensemble model. Therefore, the effectiveness of the PSO-SFLA-selected features was not judged solely by an XGBoost-like model structure.

To avoid information leakage, the three feature selection methods were applied only to the training set, and the selected feature subsets were then used to train the prediction models and evaluate their performance on the independent testing set.

2.5.3. Model Introduction and Construction

To address the spatial variability of coastal soil salinity and the nonlinear characteristics of spectral signals, an ensemble-learning framework was developed instead of using a single model. Instead, an ensemble inversion framework centered on an ensemble learning strategy was constructed [24]. The ensemble architecture was designed to improve predictive performance and model reliability through multi-level learning under the current experimental conditions. A two-stage structure is adopted by this model: the first stage (primary base learning layer) is responsible for deeply extracting patterns of various linear and nonlinear dimensions from the optimized low-dimensional features; the second stage (secondary meta-learning layer) focuses on correcting systematic residual errors and biases in the primary outputs, thereby achieving deep complementarity between physical mechanisms and fitting capabilities [12]. The selection of base learners is important because the diversity of the underlying physical drivers of the algorithms and the differences in internal structures must be balanced to ensure that predictive feature descriptions with multimodal perspectives can be received by the meta-learner. In this study, Ridge Regression (Ridge), Random Forest (RF), and Extra Trees (ET) were introduced as the primary processing engines, while Extreme Gradient Boosting (XGBoost) was selected as the secondary meta-learner.

Ridge regression compresses the regression coefficients associated with collinear and redundant bands by introducing an L2 penalty term into the conventional ordinary least squares (OLS) framework [38]. This not only guarantees that the entire inversion system does not fall into extreme noise traps, but more importantly, provides a stable global linear trend line for the model. During node splitting, a mechanism for discovering optimal splitting fields is possessed by RF, which can acutely capture the nonlinear aggregation characteristics between specific spectral index thresholds and weak salinity variations. Furthermore, the randomness of decision factor selection is maximized by ET [39]; compared to RF, strict local purity optimums are discarded by ET, endowing it with a stronger global smoothing capability. The combination of RF and ET improved the prediction stability for extremely high- and low-salinity samples. These three models—encompassing linear methods, ensemble trees, and randomized trees—form a heterogeneous ensemble, complementing each other’s capabilities in extracting diverse data features. Based on a strict second-order Taylor expansion, the loss optimization function is approximated and solved by XGBoost, which directly introduces an explicit regularization term (constraining tree depth and leaf weights) from within the objective function to control model complexity [40]. It assumes the responsibility of receiving the global linear fitting from Ridge, the fine local clustering from RF, and the smoothing generalization information from ET, while continuously performing iterative compensation in the descent direction of the negative gradient of the residuals from the previous round. After the outputs of the three base learners are integrated at the XGBoost layer, a more reliable and accurate salinity prediction can be obtained for the current dataset.

To avoid information leakage and reduce overfitting in the ensemble, the meta-features used for training the XGBoost meta-learner were generated using a strict out-of-fold prediction strategy with 5-fold cross-validation. Specifically, the training set was divided into five folds; in each iteration, four folds were used to train the three base learners (Ridge, RF, and ET), and the remaining fold was used to generate validation predictions. This process was repeated five times until each training sample obtained one out-of-fold prediction from each base learner. The three out-of-fold prediction vectors were then concatenated to form a three-dimensional meta-feature matrix for training the XGBoost meta-learner. For the independent test set, each base learner was retrained using the full training set, and its predictions on the test set were used as meta-features for the already trained XGBoost meta-learner. This ensured that the meta-learner was trained only on predictions generated from samples not used in the corresponding base-model fitting process, thereby effectively reducing self-training bias.

Similar hyperparameter optimization strategies have been widely used in soil salinity inversion studies based on machine learning models. For example, Liu et al. optimized key XGBoost hyperparameters, including learning rate, number of estimators, maximum tree depth, regularization terms, and subsampling ratio, for soil salinity inversion using remote sensing data [41]. To improve model reliability and ensure a fair comparison among different algorithms, hyperparameter tuning was performed within the training set using five-fold cross-validation. The independent testing set was not used during hyperparameter optimization and was retained only for final model evaluation. The hyperparameters and their search ranges are listed in

Table 2. Hyperparameters and search ranges used for model tuning.

Model	Hyperparameter	Description	Search Range
Ridge	alpha (α)	Regularization strength	[0.01, 0.1, 1.0, 5.0, 10.0]
Random Forest	n_estimators	Number of trees in the forest	[50, 100, 200]
	max_depth	Maximum depth of the tree	[None, 3, 5, 8, 10]
	min_samples_split	Minimum samples required to split	[2, 5, 10]
	min_samples_leaf	Minimum samples at a leaf node	[1, 2, 4]
Extra Trees	n_estimators	Number of trees in the forest	[50, 100, 200]
	max_depth	Maximum depth of the tree	[None, 3, 5, 8, 10]
	min_samples_split	Minimum samples required to split	[2, 5, 10]
	min_samples_leaf	Minimum samples at a leaf node	[1, 2, 4]
XGBoost	n_estimators	Number of boosting trees	[50, 100, 200, 300]
	max_depth	Maximum depth of a tree	[2, 3, 4, 6]
	learning_rate	Step size shrinkage	[0.01, 0.05, 0.1, 0.2]
	subsample	Subsample ratio of training instances	[0.6, 0.8, 1.0]
	min_child_weight	Minimum sum of instance weights needed in a child	[1, 2, 4, 6]
	reg_lambda	L2 regularization term on weights	[0.1, 1.0, 2.0, 5.0]

.

To further reduce the risk of overfitting during model construction, several model-complexity control strategies were adopted. For XGBoost, shallow regression trees were used by limiting the maximum tree depth, and subsampling of training instances was applied to reduce the dependence of the model on a limited number of input variables. Ridge regression was included as a regularized learner, in which the L2 penalty was used to shrink regression coefficients and improve model stability under potential multicollinearity among spectral variables. During both feature selection (especially PSO-SFLA) and model training, the 5-fold cross-validation strategy was consistently applied across all algorithms. The entire dataset was randomly divided into five subsets; in each iteration, one subset was used as the validation set, and the remaining four subsets were used as the training set. This repeated cross-validation process ensured overall model stability and further reduced the risk of overfitting caused by random data splitting. Simultaneously, to objectively verify the accuracy and reliability of this ensemble system, standard individual XGBoost, RF, ET, and Ridge predictors were independently trained in this study, ensuring strict consistency in physical control variables and parameter tuning scales. These will serve as baseline reference models and will be utilized for comprehensive alignment inspection and verification of downstream accuracy.

2.5.4. Model Interpretation Method Based on SHAP

To resolve the “black-box” problems associated with advanced machine learning and deep decision tree ensembles, and to extricate them from the purely data-overfitting dilemma, the SHapley Additive exPlanations (SHAP) interpretation method based on game theory was introduced in this study. Fundamentally, a TreeExplainer is employed to directly quantify the physical boundaries of nonlinear ensemble systems (such as RF, ET, and the ensemble framework). By probing the fluctuation weights and the positive/negative polarity of contributions generated by different values of each derivative input feature within the model output system, the direction and intensity of the effects of each feature variable are elucidated [42]. Consequently, the logical self-consistency of the data-driven inversion model is clarified from the perspective of agricultural remote sensing and physiological mechanisms.

2.5.5. Model Performance Evaluation

The predictive performance of the soil salinity inversion models was evaluated by comparing the measured EC values with the predicted EC values on the independent testing set. Three evaluation metrics were used in this study [16]: the coefficient of determination (R²), the standardized root mean square error (S_RMSE), and the ratio of performance to interquartile distance (RPIQ). R² was used to evaluate the degree of agreement between measured and predicted values. S_RMSE was used to describe the relative prediction error after normalization by the mean observed value. RPIQ was used to assess the predictive ability of the model relative to the interquartile range of the measured values. A higher R², a lower S_RMSE, and a higher RPIQ indicate better model performance [43]. The specific research workflow is shown in Figure 3.

RPIQ was calculated as the ratio of the interquartile range of the measured values to RMSE. Because the interquartile range is based on the central 50% of the observed data, RPIQ provides a robust evaluation of model performance when soil salinity values show skewed distributions or contain extreme observations. According to the classification criterion proposed by Ludwig et al., model performance can be classified as poor or unreliable when RPIQ < 2.02, capable of distinguishing high and low values when 2.02 ≤ RPIQ < 2.70, approximate quantitative when 2.70 ≤ RPIQ < 3.37, good quantitative when 3.37 ≤ RPIQ < 4.05, and excellent when RPIQ ≥ 4.05 [44].

The coefficient of determination R², S_RMSE, and RPIQ were calculated according to Equations (1), (3), and (4), respectively.

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

(1)

$$\mathrm{R}\mathrm{M}\mathrm{S}\mathrm{E}=\sqrt{{\displaystyle \frac{{\sum }_{\mathrm{i}=1}^{\mathrm{n}}{\left({\mathrm{y}}_{\mathrm{i}}-\widehat{{\mathrm{y}}_{\mathrm{i}}}\right)}^2}{\mathrm{n}}}}$$

(2)

$${\mathrm{S}}_{\mathrm{RMSE} }={\displaystyle \frac{\mathrm{RMSE}}{\stackrel{-}{\mathrm{y}}}}$$

(3)

$$\mathrm{RPIQ}={\displaystyle \frac{\mathrm{IQR}}{\mathrm{RMSE}}}={\displaystyle \frac{{\mathrm{Q}}_3-{\mathrm{Q}}_1}{\mathrm{RMSE}}}$$

(4)

where ${\mathrm{y}}_{\mathrm{i}}$ represents the true field-sampled value, $\widehat{{\mathrm{y}}_{\mathrm{i}}}$ denotes the model-predicted value, $\overline{\mathrm{y}}$ indicates the mean of the field-sampled values, $\mathrm{I}\mathrm{Q}\mathrm{R}$ is the interquartile range representing the spread of the core population, ${\mathrm{Q}}_1$ and ${\mathrm{Q}}_3$ signify the first (25th percentile) and the third (75th percentile) quartiles of the observed values, respectively.

3. Results

3.1. Comparative Analysis and Optimization of Multiple Spectral Feature Selection Algorithms

Within the high-dimensional spectral feature set comprising a total of 37 dimensions constructed in this study (including 5 primitive bands and 32 vegetation/salinity indices), extensive information redundancy and multicollinearity issues inevitably existed. To obtain the feature subset with the highest value for model prediction, three algorithms VIP, MultiSURF, and PSO-SFLA were independently employed for feature screening and comparative evaluation.

3.1.1. Screening Based on the Linear Projection VIP Algorithm

The Variable Importance in Projection (VIP) method focuses on evaluating the linear variance contribution of explanatory variables to the target response variable. From the generated VIP score plot, Figure 4 shows that, by utilizing 1.0 as the absolute screening threshold for feature importance, 9 out of the total 37 features (distributed in the orange region of the figure) surpassed the dashed threshold line of 1.0 and were subsequently designated as imperative features; meanwhile, the remaining features (distributed in the cyan region) were discarded as invalid background noise. The stepped distribution of the explanatory power of spectral features for salinity was clearly reflected in this figure. However, the results indicated that although a substantial amount of low-value noise bands was isolated by the VIP method to a certain extent, redundant and highly correlated indices still remained in the selected feature subset, constrained by the underlying global linear assumption of Partial Least Squares Regression (PLSR). This implies that deep multicollinearity among composite spectra in complex surface environments cannot be effectively disrupted by VIP, leading to obvious redundancy that persists in its dimensionality reduction results.

3.1.2. Screening Based on the Nearest Neighbor Distance-Driven MultiSURF Algorithm

The comprehensive distinguishing capability of features between inter-class and intra-class samples in a multidimensional space is evaluated by MultiSURF through calculating local nearest-neighbor sample class distances. Based on Figure 5, the radial annular distribution of the weights corresponding to the entire 37-dimensional variables could be intuitively observed. The inner red dashed baseline circle within the figure represents the invalid boundary determined by the model, where the weight equals zero; any inward contraction denotes redundant features that exert negative interference on the target prediction. Effective feature variables surpassing the safe threshold of the 95% confidence interval were highlighted in red and retained as the optimal feature subset. Based on an analysis of the selected list, compared to the singular linear orientation of the VIP method, the nonlinear synergistic correlations deep within the hidden layers of the data were indeed significantly captured by MultiSURF. However, it can be revealed from the dense cluster of high-priority red rays in the polar plot that, when confronting sample sets characterized by highly heterogeneous farmland salinity patches and extremely complex micro-scale soil water-salt transport, multiple cross-repetitive spectral indices reflecting highly similar physical salinity stress were inevitably preserved by MultiSURF due to its inherent lack of a global swarm intelligence evolutionary elimination mechanism.

3.1.3. Optimization Based on the Swarm Intelligence-Driven PSO-SFLA

In contrast to the limitations associated with traditional filtering-based algorithms, the advantages of overcoming the spectral multicollinearity problems induced by complex coastal surface environments were demonstrated by PSO-SFLA, which adopts a wrapper-based feature optimization strategy. Through a five-fold cross-validation process, the frequently selected feature variables were confirmed as the final optimal feature subset. In this study, the frequency threshold was set to 5 based on the stability-selection principle. Since 5-fold cross-validation was used, a frequency of 5 indicates that a feature was selected in all five validation folds, corresponding to a 100% selection frequency. This strict criterion was adopted to retain only highly stable variables and to reduce the inclusion of unstable, redundant, or noise-sensitive spectral features. In conjunction with Figure 6, a stepped distribution regarding the stability of features could be intuitively observed, whereby the sector radius in the polar coordinates directly corresponds to the selection frequency of the designated variables. Subsequent to stability validation, a massive number of bands subject to information redundancy and multicollinearity interference were intercepted within the gray and yellow-orange sectors—representing low frequencies (1 to 4 times)—and subsequently discarded. Conversely, a total of 11 feature (GNDVI, NDSI_reg, SI1, OSAVI, SI1_reg, RDVI, SI-T, NDSI, NIR, WDVI, Red) variables were consistently selected in all 5 cross-validation folds (a full frequency of 5 times). The feature selection results showed that the initial 37 candidate variables were reduced to a compact subset of 11 core variables. This reduction substantially decreased the input dimensionality before model construction and helped mitigate the overfitting risk associated with high-dimensional predictors under a limited sample size. The selected variables retained the most informative spectral, vegetation-index, and texture-related information for soil salinity prediction.

3.2. Testing and Comparative Accuracy Evaluation of Predictive Models

Utilizing the measured soil salinity as the dependent variable and the optimal feature combinations retained by the three different selection methods as independent input variables, a series of surface soil salinity inversion models were developed. Five different algorithms (Ensemble, XGBoost, Ridge, RF, and ET) were integrated into these models, and three distinct feature selection methods (VIP, MultiSURF, and PSO-SFLA) were employed. The accuracy of different salinity inversion models is shown in

Table 3. Evaluation of modeling accuracy of different salinity inversion models.

Methods	Screening	Train			Test
		R2	SRMSE	RPIQ	R2	SRMSE	RPIQ
Ensemble	VIP	0.725	0.305	3.134	0.650	0.343	2.812
	MultiSURF	0.768	0.280	3.413	0.718	0.308	3.134
	PSO-SFLA	0.812	0.252	3.791	0.758	0.285	3.382
Ridge	VIP	0.662	0.338	2.828	0.540	0.393	2.453
	MultiSURF	0.664	0.337	2.836	0.427	0.438	2.199
	PSO-SFLA	0.630	0.354	2.702	0.598	0.367	2.623
Random Forest	VIP	0.648	0.345	2.771	0.612	0.361	2.670
	MultiSURF	0.685	0.327	2.929	0.652	0.342	2.820
	PSO-SFLA	0.717	0.310	3.090	0.667	0.334	2.882
Extra Trees	VIP	0.656	0.341	2.802	0.585	0.373	2.582
	MultiSURF	0.757	0.287	3.335	0.693	0.321	3.001
	PSO-SFLA	0.779	0.273	3.496	0.703	0.316	3.052
XGBoost	VIP	0.650	0.344	2.777	0.630	0.353	2.734
	MultiSURF	0.689	0.324	2.948	0.672	0.332	2.904
	PSO-SFLA	0.729	0.303	3.159	0.676	0.330	2.922

.

From the perspective of model performance, the highest predictive accuracy across all feature selection strategies was exhibited by the nonlinear ensemble model. Under this model framework, the highest determination coefficients (R²) on the test set among all models, reaching 0.650, 0.718, and 0.758 for the VIP, MultiSURF, and PSO-SFLA methods respectively, were achieved. Simultaneously, low S_RMSE and favorable RPIQ values were maintained, indicating good predictive performance. In contrast, relatively weaker predictive efficacies were demonstrated by the Ridge, RF, and ET models. The Ridge model, in particular, performed poorly when combined with the MultiSURF method, yielding an R² of only 0.427. Analyzed from the perspective of feature selection efficacy, the VIP method could only support the test set R² to reach 0.650 within the ensemble model, with an RPIQ of merely 2.812. In the Ridge regression tests, although an R² of 0.664 on the training set was achieved when driven by MultiSURF, the R² on the validation test set was only 0.427. In contrast, because of its optimization-based feature selection strategy, higher accuracy was demonstrated by the PSO-SFLA even within the Ridge model (with an R² of 0.630 on the training set and 0.598 on the test set). For the remaining models utilizing the PSO-SFLA algorithm, the test validation set R² values generally ranged from 0.667 to 0.758. Based on the RPIQ indicators, it was also corroborated that the predictive capabilities of all models were enhanced following the application of the PSO-SFLA algorithm. Notably, the Ensemble–PSO-SFLA model achieved an RPIQ value slightly above 3.37, indicating good quantitative prediction performance according to the adopted RPIQ classification criterion. Conversely, a maximum RPIQ value of 3.134 was recorded for models based on the VIP and MultiSURF methods, implying that although their predictive power is limited, they can still be utilized for the preliminary classification of soil salinity.

At the model validation level, the training and testing results showed generally consistent trends, indicating that the adopted validation strategy effectively reduced the risk of overfitting. For example, under the PSO-SFLA feature selection strategy, the RF model achieved R² values of 0.717 and 0.667 on the training and testing sets, respectively. Similar trends were also observed for ET and XGBoost, suggesting that these models maintained stable predictive performance on the independent testing set. A cross-model comparison further showed that, when dealing with soil salinity inversion under complex and highly variable field conditions, the tree-based nonlinear models achieved higher predictive accuracy than Ridge, which was used as the linear baseline. Specifically, the maximum validation R² of Ridge was 0.598, whereas the test-set R² values of ET and XGBoost reached 0.703 and 0.676, respectively. These results indicated that the relationship between surface soil salinity and spectral features was not purely linear, but involved deeper nonlinear response patterns.

It can be distinctly observed from Figure 7 that the accuracies of all models at low-salinity points are significantly higher than those at high-salinity points. Compared with the VIP and MultiSURF feature selection methods, the scatter point distributions under PSO-SFLA aggregate much more tightly around the fitting line.

3.3. Interpretability Analysis of the Model

3.3.1. Contribution Analysis of PSO-SFLA-Selected Spectral Features

To further interpret the contribution of the spectral features selected by PSO-SFLA, SHAP analysis was performed on the final ensemble model. The analysis focused on the 11 selected variables, including GNDVI, NDSI_reg, SI1, OSAVI, SI1_reg, RDVI, SI-T, NDSI, NIR, WDVI, and Red. The contribution patterns of the 11 spectral features selected by PSO-SFLA in the ensemble model are shown in Figure 8. Among them, GNDVI showed the highest mean absolute SHAP value, indicating that it had the strongest overall influence on the model output. This result suggests that vegetation-related spectral responses played an important role in soil salinity estimation. Red and NIR also showed relatively high SHAP contributions, ranking only after GNDVI. Their high importance indicates that the final model captured not only vegetation stress information but also direct spectral responses associated with soil background reflectance, surface brightness, and salt-affected soil exposure. In addition, OSAVI, NDSI_reg, RDVI, and WDVI showed moderate contributions, further confirming that vegetation indices provided useful information for predicting soil salinity under UAV multispectral observation conditions. These indices are closely related to vegetation coverage, canopy reflectance, and growth status, which may change under salt stress.

The SHAP values also showed whether each feature increased or decreased the predicted salinity. Positive SHAP values indicate that the corresponding feature increased the predicted soil salinity, whereas negative SHAP values indicate that the feature decreased the predicted salinity. For the dominant features, especially GNDVI, Red, NIR, and OSAVI, the wide distribution of SHAP values suggests that their effects on model prediction varied among samples. This result suggests a nonlinear relationship between multispectral features and soil salinity. Instead, it may have been affected by vegetation cover, soil exposure, moisture conditions, and the spatial heterogeneity of salinity in coastal areas. Meanwhile, SI1, SI1_reg, SI-T, and NDSI showed relatively lower SHAP rankings. This does not mean that these salinity-related indices were unimportant. Rather, their independent marginal contributions in the final ensemble model were weaker than those of vegetation-related indices and original spectral bands. One possible reason is that part of their information overlapped with Red, NIR, and vegetation indices. Another possible reason is that salinity-sensitive indices may interact with vegetation status and soil background conditions in a nonlinear manner.

Overall, the SHAP results indicate that the final model did not rely on a single type of spectral variable. Instead, it integrated information from vegetation indices, salinity-sensitive indices, and original multispectral bands. The relatively high contributions of GNDVI, Red, NIR, OSAVI, and RDVI suggest that soil salinity prediction in the study area was strongly associated with both vegetation stress responses and red–near-infrared reflectance characteristics. Therefore, the SHAP analysis provides an interpretable explanation for the PSO-SFLA feature selection result and further demonstrates that the selected 11 features had clear and differentiated roles in soil salinity prediction.

3.3.2. Meta-SHAP Interpretation of Base Learners in the Ensemble Model

Since the ensemble model adopted in this study was constructed using Ridge, RF, and ET as base learners and XGBoost as the meta-learner, Meta-SHAP analysis was further conducted to reveal how the prediction outputs of the three base learners contributed to the final fused output of the top-layer XGBoost model. As shown in Figure 9, the prediction output of the RF base learner showed the largest contribution to the final ensemble prediction across all three feature selection methods. Its Meta-SHAP values exhibited the widest distribution range, indicating that RF had a strong influence on the final model output. In general, high RF prediction values tended to increase the final predicted soil salinity, as indicated by the concentration of high-value red points on the positive side of the SHAP axis. In contrast, low RF prediction values tended to reduce the final prediction, as indicated by the concentration of low-value blue points on the negative side of the SHAP axis. This result suggests that RF provided a stable and dominant nonlinear prediction signal for the meta-learner.

Under the VIP feature selection method (Figure 9a), the ET base learner showed a relatively complex contribution pattern. Low ET prediction values mainly produced negative Meta-SHAP values, whereas some high ET prediction values contributed positively to the final output. However, compared with RF, the distribution range of ET was narrower, and the separation between high-value and low-value samples was less distinct. The Ridge base learner showed no clear directional pattern under the VIP-based feature subset, because both high and low Ridge prediction values corresponded to positive and negative Meta-SHAP values. This indicates that the linear information provided by Ridge was relatively unstable when the VIP-selected features were used.

Under the MultiSURF feature selection method (Figure 9b), ET showed an inconsistent contribution pattern. Lower ET prediction values were partly associated with positive Meta-SHAP values, whereas higher ET prediction values were partly distributed on the negative side of the SHAP axis. This result suggests that the ET prediction signal was not fully aligned with the final fused output under the MultiSURF-selected feature subset. One possible reason is that the local-neighbor-based feature selection process retained some redundant or collinear variables, which may have caused divergence among the base learners. Similar to the VIP condition, the Ridge base learner still showed no stable contribution direction, indicating that its linear prediction signal contributed limited regular information to the meta-learner.

Under the PSO-SFLA feature selection method (Figure 9c), the contribution patterns of all three base learners became more consistent. RF, ET, and Ridge all showed a clearer positive association with the final ensemble output. Their high prediction values were mainly distributed on the positive side of the SHAP axis, whereas their low prediction values were mainly distributed on the negative side. This pattern indicates that the feature subset selected by PSO-SFLA improved the consistency among the base learners and reduced potential conflicts in their prediction outputs. Therefore, the PSO-SFLA-selected low-dimensional feature subset not only improved model accuracy, but also enhanced the interpretability and internal consistency of the ensemble model.

3.4. Spatial Distribution Characteristics of Surface Soil Salinity

Based on the selected PSO-SFLA feature optimization strategy combined with the ensemble inversion model, a high-resolution spatial distribution map of soil salinity was generated in this study. The spatial variability of soil salinity levels within the study area is revealed by this inversion map. It is intuitively apparent from the inversion distribution map (Figure 10) that the salinity levels of the two fields are generally situated within a low-to-medium environment, yet prominent spatial heterogeneity is also exhibited. The soil salinity across the majority of the area is maintained within a lower range (predominantly characterized by large continuously distributed green and yellow patches). However, irregular punctate and small-block accumulations of salinity (predominantly represented by orange and red) appear at the edges of the fields and in areas with uneven topography. This phenomenon, which constitutes localized high-salinity spots, is primarily associated with small-scale microtopography and lateral water flow within the fields; in slightly depressed areas, water converges and subsequently evaporates, causing free salts to easily remain in situ. Conversely, in flat areas with superior drainage, salts are less likely to accumulate excessively.

Notably, the field sampling and UAV imagery acquisition for this experiment were conducted in early December. During this period, winter wheat was in its seedling stage, and the vegetation coverage was relatively low. The shadowing effect exerted by the weak wheat seedling canopy on the surface soil was minimal, thereby enabling pure, high-quality bare-soil reflectance spectral information to be captured by the UAV multispectral imagery. The mixed-pixel interference and spectral saturation effects universally present during the vigorous crop growth periods of summer and autumn were effectively bypassed during this phase. Consequently, the original spectral data utilized for modeling, alongside salinity indices such as NDSI_reg and SI1_reg, were capable of fully exerting their salinity-characterizing functions, thereby maximizing the sensitivity of the model to the background soil salinity information. From the perspective of the laws governing soil water-salt transport, temperatures consistently decline in early winter, and both atmospheric evaporation and crop transpiration are significantly weakened. Distinct from the processes driven by high summer and autumn temperatures—where strong upward capillary rise and surface accumulation of deep salts are induced, precipitating secondary salinization—the low-temperature environment substantially decelerates the process of soil capillary water movement, stabilizing the transport of surface soil salinity. This mechanistically explains the reasons behind the low overall level of inverted salinity and the absence of extensive salt-crust formations with severe salt accumulation on the land surface.

In summary, the spatial distribution pattern delineated in this soil salinity inversion map is the comprehensive manifestation of the multifactorial coupling effects among seasonal climate, farmland vegetation dynamics, and microtopographic differences. Reliable spatial data foundations for targeted salt-leaching and amelioration treatments can be provided by these research findings.

4. Discussion

4.1. The Role of Feature Optimization and Interpretation of Physical Mechanisms

In this study, a significant and systematic divergence trend regarding inversion accuracy was exhibited by models constructed upon the identical ensemble framework but adopting varying feature screening schemes (Ensemble-PSO-SFLA > Ensemble-MultiSURF > Ensemble-VIP). It is indicated by this result that when utilizing UAV multispectral data for machine learning inversions, feature selection consistently remains the critical step determining the ultimate predictive capability of the model [45]. Although abundant spectral details are provided by remote sensing imagery, issues of feature redundancy and multicollinearity are also readily introduced [46]. If a vast array of noise-laden features is directly fed into an algorithm, not only is the computational burden immense, but under conditions of limited sample sizes, irrelevant environmental noise is highly susceptible to being excessively learned by the model rather than valid soil salinity patterns. Consequently, when confronted with unseen data, the generalization capability of the model is drastically constrained. The importance of feature selection and surface-condition differences can also be observed when the present results are compared with previous UAV-based soil salinity inversion studies. Yu et al. [47] used UAV multispectral imagery in the Yellow River Delta and constructed soil salinity retrieval models using PLSR, MLR, BPNN, SVM, and RF. Their optimal SSRI-based RF model achieved a validation R² of 0.745, an RMSE of 1.879, and an RPD of 2.211, indicating that the construction and screening of salinity-sensitive spectral information played an important role in improving inversion accuracy. Zhao et al. [48] further showed that UAV multispectral soil salinity inversion accuracy varied substantially among different surface-cover conditions. Their optimal model achieved an R² of 0.707 for bare land and 0.836 for agricultural land with vegetation cover, suggesting that vegetation coverage, soil background, and surface conditions can strongly affect model performance. In the present study, the Ensemble–PSO-SFLA model achieved an R² of 0.758, an S_RMSE of 0.285, and an RPIQ of 3.382 on the independent testing set. Therefore, the performance of the proposed model should not be interpreted simply as a direct numerical comparison of R² values, but as evidence that PSO-SFLA-based feature optimization can provide stable prediction ability under coastal saline-alkali farmland conditions with strong spatial heterogeneity.

The substantial disparity in efficacy among the three feature optimization methods adopted in this study stems precisely from the divergences in their underlying screening mechanisms. Derived from Partial Least Squares Regression, the VIP method belongs to a typical embedded evaluation paradigm, excelling at rapidly eliminated explicitly irrelevant features; however, the linear relationships among variables are primarily captured by it. Because the spectral reflectance of authentic soil is extremely intricate, and the spectral reflectance of coastal saline-alkali soil is not merely a simple linear superposition, numerous nonlinear spectral responses are frequently neglected by VIP. In contrast, acting as a filtering-based algorithm, a certain degree of nonlinear aggregation capability is captured by MultiSURF through computing nearest-neighbor sample distances; yet, highly similar, redundantly extracted indices still cannot be effectively identified and eliminated by it. The optimal-performing PSO-SFLA is classified as a wrapper-based feature optimization algorithm, whereby the cross-validation S_RMSE is directly used as the criterion for evaluating candidate feature subsets. It is ensured by this feature screening methodology that the 11 conclusively selected feature subsets constitute the optimal subset. This explanation is consistent with previous studies on feature selection and soil salinity inversion. VIP is derived from the PLSR framework and mainly identifies important variables within a linear latent-variable structure; therefore, it may be limited when soil salinity is affected by nonlinear interactions among spectral reflectance, vegetation condition, moisture, and soil background. Previous soil salinity remote-sensing studies have also reported that the relationship between spectral variables and salinity is nonlinear and controlled by multiple factors, making simple linear combinations insufficient for accurately representing salinization processes. In contrast, Relief-family algorithms such as MultiSURF can capture certain complex associations through nearest-neighbor comparisons, but they remain filter-style methods that evaluate feature relevance before model construction. Wrapper-based methods, such as PSO-SFLA, can directly optimize feature subsets according to model prediction performance. Similar conclusions were reported by Xie et al., who found that appropriate feature selection reduced input dimensionality and improved soil salinity estimation accuracy [49], and by Wang et al., who emphasized that feature selection combined with model optimization can improve model generalization and robustness in heterogeneous saline environments [50]. In the present study, PSO-SFLA reduced the original spectral feature space to 11 core variables, weakened the influence of redundant and collinear indices, and made the subsequent SHAP-based interpretation more physically meaningful. Previous SHAP-based soil salinity studies have also shown that interpretable machine learning can help link black-box model outputs with environmental mechanisms.

In addition, the interpretability and practical physical significance of the model are also facilitated by feature optimization. A highly parameterized model containing dozens or even hundreds of black-box bands is extremely difficult to be clearly explained and applied to practical agronomic decision-making. Through the precise screening executed by the PSO-SFLA, the focus of the model is strictly localized onto the core variables most sensitive to fluctuations in soil salinity and moisture. The anti-noise capability of the model is not only elevated by this dimensionality reduction, but tight linkages between the data model and the genuine physical processes such as ion absorption and water evaporation within actual agro-geological environments are inherently established.

4.2. Inversion Potential and Mechanism Analysis of the Ensemble Model

It is indicated by the data derived from this study that the most optimal predictive efficacy in coastal saline-alkali land inversion was achieved by the PSO-SFLA-driven ensemble model (independent test set R² = 0.758, S_RMSE = 0.285, RPIQ = 3.382); its accuracy is significantly superior to other model combinations and existing relevant research findings. This indicates that the proposed framework provided good quantitative prediction performance for soil salinity inversion in coastal saline-alkali farmland. The use of RPIQ further strengthens this evaluation because it is based on the interquartile range of the observed data and is therefore suitable for heterogeneous and potentially skewed soil salinity distributions [43]. The higher R² and RPIQ values obtained in this study suggest that the Ensemble–PSO-SFLA model improved prediction accuracy and reliability under the current coastal salinity conditions. This improvement can be attributed to the ability of PSO-SFLA to select stable and informative spectral features and the capacity of the ensemble framework to integrate complementary linear and nonlinear information from different base learners. Therefore, the proposed model not only improved overall prediction accuracy but also reduced prediction bias and smoothing effects for extreme high-salinity samples.

The core advantages of the ensemble model primarily stem from its robust nonlinear fitting and generalization capabilities. The spatial variability of coastal saline-alkali lands constitutes a complex process driven by the multiple coupling of multifarious natural and anthropogenic factors, and its mapping relationship with multispectral features exhibits a high degree of nonlinearity and random mutability. Any singular machine learning baseline model inevitably carries its specific inductive bias [51]. Due to this bias, the true nonlinear manifolds can only be approximated by a single model from a specific, narrow dimensionality, rendering it exceedingly difficult to simultaneously accommodate both global stable trends and local extreme-value mutations. The merit of an ensemble model resides in its capacity to integrate base learners. In this study, the Ridge model characterized by strict L2 penalty regression constraints, the nonlinear RF learner reliant on parallel voting within ensemble subspaces, and the ET evolutionary tree model which pushes splitting randomness to the extreme limit, were juxtaposed as bottom-layer observers. Because dissimilar underlying logics are employed by them to process data, a high degree of uncorrelated complementarity is manifested by the residual distributions they leave behind when facing the identical sample set [52]. A robust data foundation is thereby provided for the XGBoost aggregator to correct systematic biases. The essence of an ensemble model lies not merely in selecting a single optimal model, but rather in intelligently amalgamating multiple suboptimal yet diversified models. By integrating foundational learners equipped with disparate mechanisms, latent patterns within the data from multidimensional perspectives are capable of being captured by the model. The prediction errors of individual base learners are often uncorrelated, which constructs a more solid foundation for the meta-learner to rectify systematic biases and integrate multi-layered feature information, thereby maximizing the predictive value of the ensemble [51,52].

4.3. Feasibility of Constructing Salinity Inversion Models in Small Sample Size Scenarios

In refined regional UAV quantitative remote sensing tasks, constrained by geographical limitations, harsh field sampling environments, as well as the exorbitant costs and protracted cycles of laboratory sample measurement, how to break the objective upper limit of sample numbers and utilize a limited small sample capacity to construct inversion models with acceptable predictive performance and improved reliability under limited-sample conditions has perpetually remained a critical pain point in agricultural remote sensing [53]. Aiming at the model overfitting issues readily induced by small samples, corresponding strategies across three dimensions—sampling design, feature dimensionality reduction, and algorithm architecture—were adopted in this study, thereby verifying the feasibility of constructing robust models under small sample conditions.

First, during the data sampling and dataset partitioning phase, a sampling strategy balancing both randomness and representativeness was adopted in this study. Although the overall foundational sample size was limited, the salinity gradients ranging from non-salinized to severely salinized soils were evenly covered by the sampling points. It is indicated by the statistical analysis results (

Table 4. Statistical description of salt content data.

Data Set	Mean (μS/cm)	Standard Deviation (μS/cm)	Variable Coefficient (%)
Train set	820.85	481.62	58.67
Test set	877.80	518.79	59.10
Total sample size	838.07	490.77	58.56

) that a high degree of consistency across the mean, standard deviation, and coefficient of variation distributions was preserved between the training set and the test set. The stability of the data distribution was guaranteed by this rational macroscopic sample structure, and issues of model evaluation failure caused by uneven data distribution were consequently averted.

Second, model complexity was decreased through feature dimensionality reduction. For small-sample learning, the input of excessive high-dimensional features often constitutes a primary factor leading to overfitting. In this study, the PSO-SFLA was employed to streamline the initial 37-dimensional multispectral and index features into 11-dimensional core variables. A vast amount of redundant spectral information and background noise was eliminated during this process, effectively controlling the dimensionality of the input data. Such targeted dimensionality reduction enabled the machine learning models to focus more intensely on learning physical mapping laws genuinely and strongly correlated with soil salinity, precluding their excessive learning of local random interference.

Finally, the stability of the model was elevated by exploiting the inherent advantages of the ensemble model itself. In an ensemble model, base models with distinct mechanisms are combined to operate synergistically. Outliers are subjected to fundamental constraints via linear regularization by the Ridge model, whereas Decision Trees are constructed by RF and ET through the introduction of randomized samples and feature splitting, innately conferring strong anti-overfitting capabilities upon them [54]. Furthermore, the prediction residuals of the base models are subsequently rectified by the XGBoost model. The bias and variance of the model were effectively balanced by this hybrid strategy combining a linear model, tree models, and a boosting algorithm. This is corroborated by the experimental results: under optimal configurations, the performance of the model on the training set (R² = 0.812, S_RMSE = 0.285) closely approximated that on the independent test set (R² = 0.758, S_RMSE = 0.315), demonstrating a smooth transition in accuracy. It is fundamentally substantiated that the restricted generalization capabilities resulting from limited sample sizes were effectively overcome through the aforementioned mechanisms.

4.4. Shortcoming and Prospects

Although this study has made positive progress in UAV-based soil salinity inversion in coastal saline-alkali farmland, several limitations should be acknowledged. First, concerning data timeliness and seasonal bias, the field soil samples and UAV multispectral images used in this study were collected only on 4 December 2025, during the winter wheat seedling stage. During this period, the vegetation coverage was relatively low and most of the soil surface remained exposed, which was beneficial for capturing soil background reflectance information. However, soil salinity in coastal farmland is strongly affected by seasonal variations in precipitation, evaporation, groundwater depth, irrigation, and crop growth conditions. Therefore, the applicability and robustness of the model during other key periods, such as the spring irrigation period, summer crop growth period, and autumn fallow period, still require further validation and assessment.

Second, regarding sample size and spatial representativeness, although the 90 soil samples collected using a grid-based random sampling strategy were sufficient for basic model construction and accuracy evaluation, the sample size was still relatively limited for fully characterizing the complex spatial heterogeneity of soil salinity in coastal saline-alkali farmland. In particular, the prediction accuracy at high-salinity points was lower than that at low-salinity points, indicating that the model may still have uncertainty when estimating extreme salinity conditions. Therefore, caution should be exercised when directly extrapolating the proposed model to larger coastal saline-alkali regions or areas with different soil, hydrological, and management conditions.

Future research will focus on several key directions. First, multi-season and multi-year UAV observations should be conducted to reveal the temporal dynamics of soil salinity and improve the seasonal robustness of the inversion model. Second, larger and more representative field datasets should be collected across different coastal saline-alkali regions to strengthen the spatial transferability and external validation of the model. Third, multi-source data fusion should be further explored by integrating UAV multispectral imagery with thermal infrared data, LiDAR-derived terrain factors, soil moisture information, groundwater indicators, and meteorological variables. Finally, future studies should further improve model lightweighting, cross-regional validation, and mechanistic interpretability, so that UAV-based soil salinity inversion can be more effectively applied to precision agriculture and dynamic monitoring of coastal saline-alkali land.

5. Conclusions

(1)

The quantity of feature dimensions was effectively reduced by VIP, MultiSURF, and PSO-SFLA alike. However, the most superior performance in this study was exhibited by the PSO-SFLA. The inputs were successfully compressed into 11 feature variables by it, which directly facilitated the enhancement of model construction accuracy.

(2)

Regarding model development, outstanding performance was demonstrated by the ensemble model. The Ensemble–PSO-SFLA model achieved better test-set performance, with R² = 0.758, S_RMSE = 0.285, and RPIQ = 3.382. These results indicate that the framework can effectively capture nonlinear relationships between UAV multispectral features and soil salinity within the current study area, while maintaining relatively stable prediction performance under the adopted validation strategy.

(3)

Relying on the aforementioned high-precision ensemble model, the generated spatial distribution map of soil salinity intuitively and meticulously delineated the aggregation and distribution characteristics of patches with varying salinity grades within the experimental farmlands. The heterogeneity of field salinity distribution was maximally reconstructed by these mapping results, thereby providing reliable data support for regional anti-salinization engineering and dynamic agricultural monitoring.