Soil salinity prediction using a machine learning approach through hyperspectral satellite image

Abstract

A major environmental threat is soil salinity caused by natural and human-induced processes. Therefore, soil salinity status monitoring is required to ensure sustainable land use and management. Hyperspectral satellite images can make a significant contribution to the detection of soil salinity. The increase in production in semi-arid and arid regions such as Zaghouan in the northeast of Tunisia requires good soil management because this resource is a determining factor for agricultural production. This paper aims to predict soil salinity in this area using spectral signature and features vector of the Hyperion hyperspectral image. The AutoEncoder (AE) is one of neural network architectures that were adopted for feature representation. Support Vector Machines (SVM), K-Nearest-Neighbors (KNN) and Decision Tree (DT) were used for the classification. Results showed that the AE-SVM combination outperforms among the three other approaches used for soil salinity prediction.