Complémentarité des images optiques SENTINEL-2 avec les images radar SENTINEL-1 et ALOS-PALSAR-2 pour la cartographie de la couverture végétale : application à une aire protégée et ses environs au Nord-Ouest du Maroc via trois algorithmes d’apprentissage automatique.

In this article, we evaluate the classification performance of three non-parametric algorithms (kNN, RF and SVM), using multi-temporal data from three satellites (Sentinel-1, Alos-Palsar-2 and Sentinel-2) and their combinations. The study area selected is characterized by a subhumid Mediterranean climate and a very rough topography, making it especially difficult to classify land cover. In addition, it contains a protected area named Jbel Moussa and presents exceptional biological diversity.  We have acquired and pre-processed satellite images for the period from January 1 to December 31, 2017, to track vegetation cover. Then to produce 12 scenarios, we combined the three satellites. Classification maps illustrate our approach. A total of thirty-six classifications were carried out, based on seven classes: Water, Building and Infrastructure, Bare Soil, Sparse Vegetation, Grasslands, Sparse Forest and Dense Forest. The results showed that the highest overall accuracy was provided for all scenarios by RF (53.03%-93.06%), followed by kNN (49.16%-89.63%), while the lowest classification accuracy was created by SVM (47.86%-86.08%). The study also showed a similarity between the performance of the three satellites combination and that of Sentinel-2 alone. Area estimates vary from 0.85 km2 (0.11% of the study area) to 326.84 km2 (41.31% of the Study Area) for the various classes.