Abstract
Soil moisture (SM) is a significant parameter influencing various environmental processes in hydrology, ecology, and climatology. SAR-derived remote sensing products can be valuable input features for estimating SM. In the past, the results often lacked a sufficient spatial resolution for a local application. With the new Sentinel-1 sensor it seems possible to derive more detailed SM-maps. Therefore, we utilized this sensor to test the applicability of a support vector regression (SVR) based method for SM retrieval for two different grassland-covered floodplains in north-east Germany. The model was then tested for its transferability. Moreover, it was operating exclusively with free and publicly available data. In situ data of volumetric SM were collected in 2015 for both study areas at the Elbe and Peene rivers. Remote sensing input data were VV and VH backscatter, and local incidence angle derived from Sentinel-1 images, as well as height, slope, and aspect derived from SRTM images. Additionally, the Landsat 8 NDVI product was included to compensate vegetation influences on SAR backscatter. Overall, the SVR is capable of estimating SM reasonably accurate (RMSE 9.7 and 13.8 Vol% for the individual study sites). Nevertheless, the performance highly depends on the in situ data—particularly, on the amount of samples (here 98 Peene and 71 Elbe) and the value range (10–100 Vol% Peene and 12–87 Vol% Elbe). Furthermore, the optimal input feature combination and the best performance differed between study sites. In summary, backscatter, elevation, and NDVI were the most important features for SM prediction. Models using only Radar-derived features where only 1.05 or 2.01 Vol% worse than models including optical data and are, therefore, able to estimate SM. Combining samples from both study sites slightly impaired the model. Due to varying site-conditions in terms of humidity and vegetation cover, the transferability of the SVR models is not possible for the studied sites.