Introduction Soil is a fundamental non-living component of ecosystems and plays a crucial role in agricultural production, natural resource sustainability, and environmental protection. As a medium for plant growth, nutrient cycling, and biological activity, healthy soil forms the basis of food security and ecosystem stability. Among the various forms of land degradation, soil salinization is considered one of the most severe environmental challenges, particularly in arid and semi-arid regions. Soil salinity significantly reduces soil fertility, limits plant growth, decreases biological productivity, and threatens the sustainability of rangeland ecosystems. Given the extensive distribution of arid and semi-arid lands in Iran, accurate monitoring and mapping of soil salinity are essential for sustainable land management. Conventional field-based methods, although accurate, are labor-intensive, expensive, and incapable of providing adequate spatial coverage over large areas. Furthermore, field sampling offers only point-based information and a limited representation of salinity conditions. In contrast, remote sensing technologies integrated with machine learning techniques provide efficient and cost-effective tools for large-scale soil salinity assessment and mapping. The present study aimed to evaluate the effectiveness of remote sensing-derived variables and different machine learning algorithms for soil salinity classification and mapping in the rangelands of Ramhormoz County, located in Khuzestan Province, southwestern Iran.
Materials and Methods A total of 80 surface soil samples were randomly collected from the rangelands of Ramhormoz County at a depth of 0–30 cm, with a relatively uniform spatial distribution across the study area. The geographic coordinates of all sampling locations were recorded using a Global Positioning System (GPS) receiver to ensure accurate correspondence between field observations and satellite image pixels. Soil salinity was determined by measuring the electrical conductivity of the saturated paste extract (ECe) in the laboratory. Landsat 9 imagery acquired in April 2024, close to the sampling period, was processed using Google Earth Engine (GEE). Spectral bands, topographic information (elevation), and several salinity-related spectral indices, including the Normalized Difference Vegetation Index (NDVI), Normalized Difference Water Index (NDWI), Brightness Index (BI), Salinity Index (SI), Salinity Index 1 (SI1), Salinity Index 2 (SI2), and Normalized Difference Salinity Index (NDSI), were extracted and used as predictor variables. To reduce multicollinearity among input variables, Pearson correlation analysis was conducted. Variables exhibiting strong pairwise correlations (r > 0.90) were excluded from subsequent analyses. Soil salinity classification was performed using four machine learning algorithms: Classification and Regression Trees (CART), Random Forest (RF), Support Vector Machine (SVM), and Multilayer Perceptron Neural Network (MLP). The performance of the models was evaluated using 10-fold cross-validation (K = 10), overall accuracy, weighted F1-score, and the coefficient of determination (R²).
Results and Discussion The measured soil salinity values ranged from 1.79 to 11.30 dS m⁻¹, indicating moderate variability in salinity conditions and the absence of extremely saline soils within the study area. This range provided an appropriate dataset for training and evaluating the machine learning models. Comparative analysis of the tested algorithms revealed notable differences in predictive performance. The CART model demonstrated the highest classification accuracy, achieving an overall accuracy of 75% and a weighted F1-score of 74.8%, indicating its superior capability in distinguishing salinity classes. The SVM model with a linear kernel ranked second, with an overall accuracy of 63%. In contrast, the Random Forest model, SVM models employing radial basis function (RBF) and polynomial kernels, and the MLP neural network exhibited relatively lower classification performance. The superior performance of the CART algorithm may be attributed to its simplicity, robustness, and ability to handle nonlinear relationships efficiently, particularly when working with relatively small datasets. Furthermore, CART provides highly interpretable decision rules, making it suitable for environmental applications where model transparency is desirable. The final soil salinity map indicated that approximately 13.5% of the rangelands in Ramhormoz County are affected by high salinity levels. Areas with elevated salinity were primarily concentrated in the northern, southern, and western parts of the study area. These spatial patterns were consistent with field observations, supporting the reliability of the generated salinity map and the adopted methodological framework.
Conclusion The results of this study demonstrate that integrating Landsat 9 remote sensing data with machine learning algorithms can provide an effective approach for soil salinity assessment and mapping in arid and semi-arid rangelands. Among the evaluated algorithms, the CART model exhibited the best overall performance and proved to be a reliable tool for soil salinity classification under the environmental conditions of the study area. The findings highlight the importance of selecting appropriate machine learning models when developing soil salinity maps and suggest that decision tree-based approaches may outperform more complex algorithms when sample sizes are limited. Future research should investigate the integration of higher spatial resolution optical and radar imagery, larger field datasets, and advanced deep learning techniques to further improve the accuracy and robustness of soil salinity mapping in rangeland ecosystems.