Machine Learning based Regression Model for Prediction of Soil Surface Humidity over Moderately Vegetated Fields

Abstract

The soil surface humidity parameter over vegetated fields is of great importance for controlling water consumption; prevention of salinity caused by over-irrigation; efficient use of irrigation system and improving the yield and quality of the cultivated crop. However, determination of the soil surface humidity is very difficult on vegetated fields. In order to overcome this problem, polarimetric decomposition models and machine learning based regression model were implemented. The main purpose of this study is to predict soil surface humidity on moderately vegetated fields. Thus, the study is conducted in agricultural fields of Dicle University and it consists of several stages. In the first stage, a Radarsat-2 data was obtained in 3 March 2016 and the local humidity samples were measured simultaneously with the Radarsat-2 acquisition. In the second stage, 10 polarimetric features were obtained from each cell (2x2 pixels) of ground sample by utilizing standard intensity-phase technique as well as Freeman-Durden and H/A/$\alpha$ polarimetric decomposition models. This step is repeated for all ground samples and as a result, a dataset with 156x10 lengths is formed. In the next stage, Extreme Learning Machine based Regression (ELM-R) model was used for predicting the soil surface humidity with the aid of polarimetric SAR features. For the validation of the proposed system, leave-one-out cross-validation method was applied and finally, 2.19% Root Mean Square Error (RMSE) were computed.