Optimization of land cover mapping through improvements in Sentinel-1 and Sentinel-2 image dimensionality and data mining feature selection for hydrological modeling

The characterization of land cover in a watershed has a remarkable influence on the water balance and, is a key factor in assessing complex hydrological models at the regional scale. Here, a method is proposed to generate accurate seasonal ad hoc land cover maps from a hydrological standpoint based on the runoff generation capabilities of the land cover types using remote sensing techniques with Sentinel-1 and Sentinel-2 data and a random forest approach. A multidate study is proposed, minimizing the images as much as possible: one single date for Sentinel-2 data and two dates for Sentinel-1 data. Then, the dimensionality of the satellite data is improved with texture metrics and derived spectral indices, after which a data mining feature selection is conducted through optimization of the classification algorithm, ensuring the accuracy of the final maps. The overall accuracies are remarkably high (93.29%) for the test dataset and still outstanding (85.24%) for the validation dataset. The texture metrics are the most important classification parameters and are mainly derived from VIS (B2 and B3), NIR (B5 and B6) and SWIR (B11). The results outperformed previous works that used large temporal image series and reduced the storage capacity requirements and computational time. Consequently, the hydrological response of the watersheds in terms of NRCS-CN is characterized truthfully, allowing the analysis of the potential runoff and its variation due to seasonal phenology. Future research will be focused on the analysis of rainfall-runoff models and the variability in seasonal runoff in forested watersheds in Mediterranean environments.