An Approach Based on Low Resolution Land-Cover-Maps and Domain Adaptation to Define Representative Training Sets at Large Scale

The accurate classification of remote sensing (RS) data at large scale is typically hampered by the availability of training data representative of the whole study area. To solve this problem, we propose a method that aims to enlarge existing training sets leveraging publicly available thematic products. First, the available thematic product of the target domain (DT) (RS data geographically distant from the training samples) is processed to extract few labeled target samples. These labeled target samples are jointly used with the annotated samples of the source domain (DS) (RS data where training set is available) to find a mapping space where the data are aligned. This common latent space allows us to enlarge the training set in an unsupervised (no annotated samples from the DT are required) but reliable way. The results obtained in Amazon using the Copernicus Global Land Service - Land cover (CGLS-LC) map demonstrate the effectiveness of the method. The enlarged training set achieves an Overall Accuracy (OA) of 87% compared to 80% obtained with the initial training set.