Remote Sensing and Machine Learning for Identification of Salt-affected Soils

Abstract

Approximately 6.74 million hectares (Mha) are reported under salt-affected soils (SAS) in India. The accurate identification and monitoring of these soils is important for timely decision-making and assessing the success of reclamation measures. In the study, widely used three Machine Learning (ML) techniques, namely, Logistic Regression (LR), Support Vector Machine (SVM), and Random Forest (RF) were tested and compared in the identification of SAS in the Raibareli district of Uttar Pradesh, as a part of Indo-Gangetic Plains (IGP) of India by using Landsat 8 Operational Land Imager/Thermal Infrared Sensor (OLI/TIRS) data. Various salinity indices and principal components were computed through analysis of the Landsat 8 OLI/TIRS bands, and ancillary data like nearness to canals and streams were used to identify SAS. The three models were applied to object-oriented segments generated from these rasters. Total 361 segments were considered for training, and testing of RF model, of which 130 segments belonged to SAS, and remaining segments represented other features and normal soils. In the study, 70% of samples were used for training and 30% for testing. The accuracy of trained models on testing dataset was 96%, 98%, and 98% for LR, SVM, and RF, respectively. The trained model was then used for identification of SAS in the study area. The results were validated with field observations, and high-resolution Google Earth data. It was observed that the three models were able to identify 31,954, 16,679, and 14,070 ha area of SAS, respectively, in Raibareli district, IGP of India.