On a yearly basis prediction of soil water content utilizing sar data: A machine learning and feature selection approach

Abstract

Soil water content (SWC) performs an important role in many areas including agriculture, drought cases, usage of water resources, hydrology, crop diseases and aerology. However, the measurement of the SWC over large terrains with standard computational techniques is very hard. In order to overcome this situation, remote sensing tools are preferred, which can produce much more successful results in less time than standard calculation techniques. Among all remote sensing tools, synthetic aperture radar (SAR) has a significant impact on determining SWC over large terrains. The main objective of this study is to predict SWC on a yearly basis over the vegetation-covered terrains with the aid of different machine learning techniques and SAR based Radarsat-2 data, which obtained in 2015 and 2016 years.The proposed system consists of several stages, respectively. In the feature extraction stage, the backscatter coefficients of different polarizations and the parameters obtained from different models of decomposition (Freeman-Durden and H/A/α) were combined and nine polarimetric features were formed for each sample point. In the next stage, support vector regression (SVR), generalized regression neural network (GRNN) and adaptive neuro-fuzzy inference system (ANFIS) were employed for the prediction of SWC. In the last stage, a machine learning based feature selection was implemented to the obtained feature vectors for determining optimal feature sets. Finally, a feature set with 6 parameters was determined as most optimal feature set over the SWC prediction and a slightly better performance was observed thanks to this feature set compared to the other results.