The Use of Spectral Indices to Estimate Soil Surface Moisture using Machine Learning Algorithms

Detailed information about soil moisture and its spatial and temporal distribution provides opportunity for optimized land resources utilization. Our study aimed to estimate soil surface moisture through readily availabile soil parameters and spectral index obtained from Sentinel-2 sensors using two methods, artificial neural networks (ANN) and support vector regression (SVM). There were 124 soil samples collected from three regions of Iran (Tehran, Garmsar, and Lorestan). After normalizing the data, the significance of the correlation between input variables (spectral indices and basic soil properties) and output variables (surface moisture) was evaluated statistically. In the next step, the mentioned methods were used to perform a modeling process, and the results were evaluated. The results showed that the ANN method outperformed the SVM method. Based on ANN technique, the Root Mean Square Error (RMSE), Akaike Information Criterion (AIC), coefficient of determination (R2) and Relative Improvement (RI) in the training step were 0.033, -538, 0.71, 21.25, and in the testing step they were 0.410, -266, 0.69, and 16.06, respectively. Also, RMSE, AIC, R2, and RI in the SVM method in training step were respectively 0.035, -474, 0.71, and 35.16 and in testing step were respectively 0.046, 252, 0.63, and 20.21. Using the ANN method, soil color index (CI) has been shown to estimate soil moisture more accurately than other spectral indices. Therefore, the ANN method constructs a nonlinear relationship between soil surface moisture and input parameters, which enables soil moisture to be estimated with acceptable accuracy in the study area.