Mapping Land Cover from Sentinel-2A Using Support Vector Classifier and Random Forest Regressor in the Souss Basin Morocco

The semi-arid to arid area of the Souss basin is vulnerable to flood. This natural phenomenon whose intensity is becoming increasingly alarming. Indeed, the susceptibility of the basin to floods disasters is accentuated by its rapid demographic evolution, uncontrolled land cover, anthropogenic actions and other physical factors. The land cover map represents crucial information for assessing the hydro-meteorological flood hazard as the physical and environmental vulnerabilities to this phenomenon in the Souss region. Therefore, seven optical Sentinel-2A images have been used for this purpose. After a preprocessing operation, different features were extracted from the images, including spectral, morphological, and textural variables to be analyzed using the classification and regression techniques based on the Support Vector Classifier (SVC) and Random Forest Regressor (RFR) algorithms. This operation resulted in the generation of a 10 Land cover classes map and the building density estimation. The two results were merged and enhanced by introducing additional classes from exogenous data to produce a land cover map with 27 classes at 10 m resolution. After field inspection missions, the overall accuracy of this map is 91.6% with a Kappa coefficient of 89% which indicates a very good quality. This map is of great importance for many research projects and for many operational applications. Therefore, the approach developed in this study could be used to understand the current land cover and may update this information over time as an accurate means of monitoring change, which is a vital dimension for land management decision making in the region. This map was also produced under an environmental project on modeling and mapping of flood risk, to develop an integrated management action plan for the Souss basin.