SWE retrieval by exploiting COSMO-SkyMed X-band SAR imagery and ground data through a machine learning approach

Abstract

The main objective of this work is to estimate Snow Water Equivalent (SWE) by jointly exploiting the information derived from X-band Synthetic Aperture Radar (SAR) imagery acquired by the Italian Space Agency COSMO-SkyMed satellite constellation in StripMap HIMAGE mode and manual SWE ground measurements. The idea is to verify the sensitivity of the backscattering coefficient at X-band to the SWE and, by means of a Support Vector Regression (SVR) algorithm, to estimate the SWE for the South Tyrol region, north-eastern Italy. The regressor is trained by exploiting about 1,000 simulated backscattering coefficients corresponding to different snowpack conditions, obtained with a theoretical model based on the Dense Media Radiative Transfer theory - Quasi-crystalline approximation Mie scattering of Sticky spheres (DMRT-QMS). Then, the performance is evaluated on the backscattering values derived from COSMO-SkyMed satellite images and using the corresponding ground measurements of SWE as references. The results show a correlation coefficient equal to 0.6, a bias of 10.5 mm and a RMSE of 51.8 mm between estimated SWE values and ground measurements. The limited performance could be related to the DMRT-QMS theoretical model used for the simulations that results to be very sensitive to snow grain size and may have generated a training dataset only partially representative of satellite derived backscattering coefficients used for testing the algorithm.