ANN-Based Soil Moisture Retrieval over Bare and Vegetated Areas Using ERS-2 SAR Data

Active microwave remote sensing data (e.g., radar) can be used for estimation of soil moisture beneath the ground surfaces up to 0 to 5 cm depths. A number of analytical and empirical models are available that relate the synthetic aperture radar (SAR) data to the surface soil moisture. Each of these models has its own merits and demerits. In this paper, results obtained from a study on the application of artificial neural networks (ANN) for soil moisture retrieval from ERS-2 SAR data over bare and vegetated surfaces is presented. SAR images of three dates (i.e., July 28, 2003, March 29, 2004, and May 3, 2004) were acquired over a portion of the Solani river catchment in India. A feed forward back-propagation neural network was used for establishing the relationship between surface soil moisture and terrain as well as sensor variables. A maximum of seven input variables were considered for training the ANN. These are: Digital number (DN) or backscatter coefficient ( σ deg) (as the case may be) of each pixel of SAR image selected individually, incidence angle of radar beam, land cover, surface roughness height, terrain height, leaf area index, and plant water content. Several ANN experiments were conducted and the coefficient of determination ( R2 ) and root-mean-square error (RMSE) were determined between the ANN derived soil moisture and the observed soil moisture through in-situ measurements taken concurrently with the satellite pass. It has been observed that it is the backscattering coefficient σ deg rather than DN, used as one of the inputs that produced high R2 (∼0.9) and low RMSE. The results also indicate that only a few number of input variables may be sufficient to retrieve the soil moisture with high accuracies using the ANN. The leaf area index has been found to be as good as the single bulk variable representing the vegetation characteristics in the study area. In addition, the ANN technique has yielded more accurate results than the traditional statistical regression, indicating its usefulness for soil moisture estimation from microwave remote sensing data.