Estimation of soil moisture content from L- and P-band AirSAR data: A case study in Jeju, Korea

Abstract

One of the important applications of polarimetric SAR in the geohydrology and agriculture is the estimation of surface soil moisture from the polarimetric SAR data. During the PacRim AirSAR campaign in Korea, the ground truth data about soil moisture content and surface roughness characteristics were collected. We intend to retrieve the surface parameters over the bare soil from multi-polarization and multi-frequency AirSAR data. In this study, the theoretical scattering model, the IEM model is inverted by two existing algorithms—the multi-dimensional regression technique by Dawson et al. (1997) and the inversion using 3-layer artificial neural networks (ANNs) (Fung, 1994). As the first step, backscatter coefficients are calculated based on the ground truth information, and then training patterns are generated from within the valid ranges of surface parameters using the IEM model. The trained inversion models are tested to a set of AirSAR data as well as synthetic data. Root mean square (RMS) errors of estimated soil moisture from the AirSAR data are average 3.1% in the regression and 4.2% in the inversion using the ANNs. The methods to improve the inversion accuracy are investigated. First, the normalization of signal parameters reduced the number of pixels that fail to reasonable results in the regression model. Second, the use of co-polarization ratio as input units in the ANNs inversion scheme improve the soil moisture estimation, which results in an average RMS error of 2.9%.