Evaluating and comparing supervised classification algorithms With the aim of mapping soil salinity levels using Sentinel-2 satellite imageries

Soil salinity is considered to be a major cause of desertification and destruction of environmental resources in arid and semi-arid regions. Due to the importance of conserving natural resources and also the increasing trend of soil salinity during the last few years, determining the extent of salinity spread and its severity in affected areas is especially important. Using the potential of high resolution satellite imagery and remote sensing techniques is one of the most effective ways for detecting salinity in salt affected regions. Among different satellite sensors, satellites which provide large scale multispectral satellite imageries with high spatial and spectral resolution have a high potential for assessing salinization and mapping soil salinity in study regions.

Accordingly, this paper aims to map different salinity levels in an area in Kuh-Sefid district (Qom Province), which is highly affected by salinity, using Sentinel-2 recent imageries. For this purpose, field study was conducted and salinity level was measured for several soil samples randomly collected from the site. Different salinity indicators, like salinity and vegetation indices, LST, and Digital Elevation Model of the site produced based on SRTM elevation data were also extracted from corresponding satellite images. These indicators were then used for mapping salinity levels in Kuh-Sefid district. Principal Component Analysis was used to gain the largest amount of available information and reduce the dimensionality of data cube. Based on the performed analysis, different supervised classification algorithms were used to map salinity levels and divide the site into five distinct salinity classes - normal, slightly saline, moderately saline, highly saline, and extremely saline. 

Data was analyzed based on five supervised classification algorithms, including Minimum Distance, Mahalanobis, Parallelepiped, Maximum Likelihood, and Support Vector Machine (SVM). Results indicated that the best accuracy in mapping salinity classes was obtained from SVM classifier, with overall accuracy of 92.218 and Kappa coefficient of 0.894. The results also revealed that Maximum Likelihood Classifier with overall accuracy of 90.718 has a high potential for discriminating saline surfaces and producing salinity map. In addition, more than 62% of soil types in this region are categorized in moderate, high and extreme salinity classes, which indicates that the area is highly at risk. 

Evaluating the results of salinity classes shows that the eastern areas of Kuh-Sefid are relatively more severely affected by salinity. This is due to vicinity of Qom Salt Lake and drawing of saline soil into surrounding areas. If this process continues, it will lead to loss of soil fertility and crop productivity in this region over the next few years. Due to their potential in detecting soil salinity and providing large-scale maps of salinity levels, multispectral Sentinel-2 imageries are considered to be a powerful tool in soil reclamation programs and land management studies.