Digital Mapping of Soil Salinity in the Southern Steppe Zone of Russia Based on Artificial Neural Networks and Linear Regression

Abstract Remote sensing data are an important source of information for monitoring and mapping vegetation cover. Machine-learning methods are a modern and powerful tool for data processing. However, machine-learning methods combined with remote sensing data have hardly been used for soil salinity assessment and mapping in the southern steppe zone of Russia. This paper examines the possibility of applying different spectral characteristics to map soil salinization in solonetzic complexes in the southern steppe zone of Russia (Republic of Kalmykia) using machine-learning methods. A number of predictors were considered, including reflectance coefficients in blue, green, red, and infrared spectral zones; vegetation indices (NDVI, NDVI t , TVI, SAVI, MSAVI, EVI 1 –EVI 4 ); salinity indices (SI 1 –SI 6 ); intensity indices (Int 1 , Int 2 ); brightness index (BI); and an index proposed by the authors. High-resolution images from the QuickBird (2.4 m) and SuperView-1 (2 m) satellites were used. Soil salinity was assessed using two indicators: specific electrical conductivity in water suspension (EC 1:5 ) and sodium activity (aNa 1 : 5 ). Two different machine-learning models were applied in the study: linear regression and neural networks. According to the results obtained, the linear regression model for EC 1 : 5 in 0- to 30-, 0- to 50-, and 0- to 100-cm layers has coefficients of determination ( R 2 ) of 0.53, 0.59, and 0.79 on the training sample; the test sample managed to obtain coefficients of determination of 0.49, 0.58, and 0.70, respectively. The neural-network model has significantly higher coefficients of determination: R 2 for EC 1 : 5 in 0- to 30-, 0- to 50-, and 0- to 100-cm layers on the training sample is equal to 0.68, 0.91, and 0.97, and on the test sample, 0.87, 0.86, and 0.88, respectively. This fact indicates a greater potential of this model for cartographic modeling of soil salinity. The best predictors were the following indices: NDVI t , TVI, EVI 1 , and Int 1 . The study has shown the potential of using the neural-network model and spectral indices obtained with SuperView-1 images for soil salinity mapping of solonetzic complexes in the south of the steppe zone of Russia.