ML based approach for inverting penetration depth of SAR signals over large desert areas

Abstract

Penetration depth of synthetic aperture radar (SAR) signals over a desert is a key parameter to understand the internal properties of the desert. Existing approaches for obtaining the penetration depth require good quality interferometric SAR (InSAR) data of very short temporal and long spatial baselines. Such data are often difficult to obtain in a highly dynamic desert. We propose a new machine learning (ML) based approach for inverting penetration depth of SAR signals over large desert areas by jointly using InSAR, polarimetric SAR (PolSAR) and optical remote sensing data. First, SAR scattering parameters and terrain properties are determined based on PolSAR and Landsat 5 TM multispectral data and a DEM. The penetration depth of SAR signals over a small desert area is obtained based on methods such as using a scattering model. A random forest model is then used to establish the relationship between the SAR scattering parameters and site features, and the penetration depth, and then is used to derive the penetration depth over a large desert area. The approach is applied to calculate the penetration depth of ALOS-1 PALSAR L-band signals for a large part of the Kufra Basin, an area of about 60, 000 km2. The penetration depths of four types of typical landforms in area (i.e., sandy plains, paleochannels, rocks and man-made features) are discussed in relation to the geological and climatic conditions. The average signal penetration depths over the paleochannels, sandy plains, and rocks and man-made features are 2.84 m, 1.97 m, 1.21 m, respectively. It is found that the backscattering coefficient, dielectric constant, surface roughness and mineral composition are the most important parameters in determining the signal penetration depths. An interesting point is that the existence of hematite in the sand can increase the dielectric dissipation of the sand medium and shorten the signal penetration depth.