Soil moisture estimation from microwave remote sensing data with nonlinear machine learning techniques

Abstract

Soil moisture estimation is one of the most challenging problems in the context of biophysical parameter estimation from remotely sensed data. Different approaches have been proposed in the literature, but in the last years there is a growing interest in the use of non-linear machine learning estimation techniques. This paper presents an experimental analysis in which two non-linear machine learning techniques, the well known and commonly adopted MultiLayer Perceptron neural network and the more recent Support Vector Regression, are applied to solve the problem of soil moisture retrieval from active and passive microwave data. Thank to the use of both simulated and real in situ data, it was possible to investigate the effectiveness of both techniques in different operative scenarios, including the situation of limited availability of training samples which is typical in real estimation problems. Moreover, for each scenario, different configurations of the input channels (polarization, acquisition frequency and angle) have been considered. The comparison between the two methods has been carried out in terms of different figure of merits, including error measurements and correlation coefficients between estimated and true values of the desired biophysical parameter. The results achieved indicate the Support Vector Regression as an effective alternative to the neural network approach, due to a general better estimation accuracy and a higher robustness to outliers, especially in case of limited availability of samples.