Evaluation of the Potential for Estimating Backscattering Coefficients over Bare Agricultural Soils at the Intra-Plot Scale

The objective of this study is to model backscattering coefficients over bare soils at intra-plot spatial scales (from almost 80 to 2800 m2), in a context where the plot is the reference spatial scale in most past studies. A statistical modeling approach, based on a random forest algorithm, is proposed to overcome the limits of semi-empirical or physical models pointed out in the literature and to reduce discrepancies observed between the satellite-derived backscattering coefficients and the predicted values. The experimental device was set up on a network of agricultural plots located in southwestern France during the Multispectral Crop Monitoring (MCM) experiment. The dataset combines high spatial resolution satellite images (acquired by TerraSAR-X and Radarsat-2) together with synchronous geo-located measurements of key soil parameters (i.e., top soil moisture, surface roughness, and soil texture) on consistent spatial areas. Backscattering coefficients are estimated at six intra-plot spatial scales (from ~80 to ~2800 m2), showing an exponential increase in modeling performance, and reaching higher levels of accuracy than previous work performed at the plot spatial scale (i.e., 50% of variance explained in the literature, in the best cases). The increase in signal quality with the spatial scale mainly explains the higher performance observed in the 2800 m2 area, with a correlation of 0.91 and RMSE of 0.83 dB in the X-band (for backscattering coefficients acquired with the HH polarization state). In the C-band, the values of correlation range from 0.74 to 0.80, and the RMSE from 1.65 to 1.85 dB (depending on the considered polarization state). The results also showed that the developed statistical algorithm is mainly influenced by the surface roughness and the top soil moisture, as semi-empirical or physical-based models. Soil texture does not significantly affect the algorithm.