Automatic Extraction of Weak Labeled Samples From Existing Thematic Products For Training Convolutional Neural Networks

The accuracy in classification of remote sensing (RS) images using deep learning architectures is affected by the lack of large sets of training samples. Although a significant effort is currently devoted to generate databases of annotated satellite images, these datasets may not be large enough to accurately model at global level different types of land-cover surfaces. To solve such a problem, this paper presents an unsupervised approach which aims to exploit the RS image that has to be classified and publicly available thematic products to generate a training database of weak samples representative of the considered study area. First, we harmonize the thematic map and the RS image. Then, samples having the highest probability to be correctly associated to their labels are extracted from the map by exploiting the information provided by the RS image to be classified. Finally, the weak labeled samples are used to train a convolutional neural network (CNN). Experimental results obtained training a CNN on Sentinel 2 images with weak labels extracted from the 2018 corine land cover (CLC) map demonstrate the effectiveness of the proposed method.