On The Use of Machine Learning and Polarimetry For Estimating Soil Moisture From Radarsat Imagery Over Italian And Canadian Test Sites

This research aimed at exploiting the joint use of machine learning and polarimetry for improving the retrieval of surface soil moisture (SMC) from SAR acquisitions at C- and X-band.The study was conducted on an alpine test area in Italy and two agricultural areas in Canada, for which series of Radarsat-2 (RS2) and COSMO-SkyMed (CSK) images were available along with direct measurements of SMC from in-situ stations. The analysis confirmed the sensitivity of SAR backscattering (σ°) from both sensors to the SMC variations, with similar correlations (R ≃0.5). The comparison of SMC with the Compact Polarimetric (CP) parameters, computed from the RS2 acquisitions by Radarsat Constellation Mission (RCM) data simulator pointed out that the right and left polarized signals and the Shannon entropy intensity also have some sensitivity to SMC variations, with R ≃0.4 for all the three parameters.Based on these results, two different machine learning (ML) algorithms, namely Support Vector Regression (SVR) and Artificial Neural Network (ANN) have been implemented and tested on the available data. On the South Tyrol test area, both SVR and ANN tested with different combinations of RS2 and CSK data were able to retrieve SMC with a RMSE between 4% and 6% of SMC and R between 0.78 and 0.88, depending on the combination of inputs. The ANN algorithm based on CP data was tested on the Canada areas, being able to estimate SMC with a RMSE between 2% and 5% of SMC and R between 0.85 and 0.96.