COMPARATIVE STUDY OF MACHINE-LEARNING-BASED CLASSIFIERS FOR SOIL SALINIZATION PREDICTION USING SENTINEL-1 SAR AND SENTINEL-2 MSI DATA

Under arid and semi-arid climatic conditions, salinization, naturally induced by parental materials weathering or artificially caused by anthropogenic actions, depicts a severe land degradation form, which significantly impacts crop yield in arable lands affecting food security on local and regional scales. At the same time, its adverse effects on urban infrastructures extend to land subsidence, erosion, and groundwater quality deterioration. Hence, soil salinity and sodicity monitoring using remote sensing and GIS tools has been an attractive field of research for soil scientists and experts worldwide, mainly for promoting sustainable agriculture. In this context, using Sentinel-l SAR and Sentinel-2 MSI data acquired in September-October 2019 and soil salinity data provided by the Research Institute of Soil Sciences and Agricultural Chemistry (RISSAC), four machine learning classifiers, namely, random forest (RF), maximum likelihood (ML), minimum distance (MD), and nearest neighbors (KNN) were applied and compared in terms of classification accuracy. The results revealed that RF had outperformed other classifiers, with the highest overall accuracy of 97 % and a mean percentage error (MPE) of 0.028, followed by KNN (= 96 %, 0.041), ML (= 94.5 %, 0.054), and MD (= 86.5 %, 0.135). The importance score report based on the RF classifier showed that the VV backscatter coefficient, SWIRl, SWIR2, and VH backscatter coefficient are the most important features retrieved from remotely sensed data to estimate salinization risks. Despite its strong dependence on training data, random forest depicts the most suitable algorithm for salinization detection using radar and optical data. This study can provide a timeless reference for selecting efficient machine learning algorithms in future studies to predict salinization using data retrieved from multiple sensors.