Spatial Prediction of Soil Salinity by Using Remote Sensing and Data Mining Algorithms at Watershed Scale, Northwest Iran

Abstract

Soil salinity plays an important role in agriculture production and land degradation, especially in semi-arid and arid regions. Accurate prediction of soil salinity requires evaluating crop yield, native vegetation situations, and irrigation command area management. In this study, MLR (multiple linear regression), SVMs (support vector machines) and ANNs (artificial neural networks) models were employed by using Landsat-8 OLI and GIS (Geographical Information Systems) techniques for predicting soil salinity in northwest Iran. Soil salinity was measured at 92 points (in a depth of 0–20 cm). The vegetation and soil salinity spectral indices, extracted from Landsat-8 OLI, were employed as input data. The results of this study indicated that the best-developed model for predicting soil salinity was the SVM-based model with R² (0.874) and RPD (2.32) and the lowest RMSE (11.20 dS m⁻¹). Moreover, the performance of developed models under different vegetation coverage showed that the SVM-based model yielded the best result. It was concluded that the SVM-based model is reliable for quantifying soil salinization.