River salinity mapping through machine learning and statistical modeling using Landsat 8 OLI imagery

Abstract

This study uses Landsat 8 OLI imagery and 102 in situ salinity data points to investigate salinity mapping in the Karun River, southwestern Iran. A total of 24 features, including salinity indices and Landsat 8 OLI spectral bands, were assessed using the Random Forest Feature Importance Score (RFFIS), Sobol’ sensitivity analysis, and correlation with salinity to identify the most sensitive features for salinity estimation. These included the Red and Green bands, Salinity index 2–6, Normalized Suspended Material Index (NSMI), and Enhanced Green Ratio Index (EGRI). A total of 24 regression models, including statistical, kernel-based, Neural Network (NN)-based, and Decision Tree (DT)-based models, were evaluated using statistical error metrics and global, as well as local, Moran’s I measures of residual spatial autocorrelation. The DT-based models, specifically Gradient Boosted DT (GBDT), outperformed other models, demonstrating low errors, bias, and non-significant residual spatial autocorrelation. Kernel-based models performed better than conventional linear models, while NN models tended to underfit. Residual spatial autocorrelation analysis indicated that models incorporating spatial information reduced residual autocorrelation. Landsat 8 OLI imagery effectively mapped salinity dynamics, revealing increased salinity from Gotvand to Ahvaz city due to agricultural activities and the Gachsaran formation within the reservoir.