Temporal remote sensing based soil salinity mapping in Indo-Gangetic plain employing machine-learning techniques

Abstract

Soil salinization is one of the most active land degradation processes, affecting predominantly arid, semi-arid, and dry sub-humid regions and leading to decreased agricultural yields. The Indo-Gangetic plain, which includes the irrigated command areas with arid and semi-arid climatic conditions are severely affected by secondary soil salinization. Assessing the spatial and temporal extent as well as the severity of salinization is an important step for adoption of proper reclamation measures to boost agricultural productivity in the salt affected areas. The study was conducted with this background to evaluate the extent and severity of soil salinization in alluvial plains of Mathura district of Uttar Pradesh, India. In this study, the satellite data of 6 months from January 2019 to June 2019 were pre-processed and various spectral indices were generated in Google Earth Engine. Remote sensing techniques provides an ideal platform for addressing this problem at larger scales and thus we employed Sentinel-2 derived vegetation and salinity spectral indices for distinguishing temporal change in severity of soil salinization and map the salinity as a function of these indices for the entire study area. The time series salinity analysis showed that among the various spectral indices Ratio Vegetation Index (RVI), Normalized Difference Vegetation Index (NDVI) and Normalized Difference Soil Index (NDSI) had a clear differentiation between slight, moderate and severe salinity class in the first three months of the study period and the Salinity Index II (SI_II) could differentiate for the first four months. Further, two machine learning algorithms namely Random Forest (RF) and Support Vector Machine (SVM), were used to create soil salinity prediction models making use of the soil Electrical Conductivity (EC) values of 115 ground-sampling sites as the predictand variable and the optimal spectral indices as the predictor variables. Further, we evaluated the prediction ability of different models using 12 and 24 variables combination using R2 and RMSE values. The prediction accuracy of the RF model was found to be slightly higher than that of the SVM model, and the spatial distribution pattern of soil salinity predicted by the two models were comparable. We concluded that spectral indices combined with machine learning techniques have the potential for low cost reliable spatial and temporal soil salinity distribution mapping for planning and implementation of salinity reclamation measures.