Unsupervised Multiclass Change Detection and Mapping Using Deep Neural Network

This work deals with the task of land use and land cover (LULC) change detection using multi-temporal multispectral remote sensing images. In the last few years, deep learning-based change detection methods have been successfully implemented for automatic LULC change detection. Although these elaborated methodologies achieved a very high score in the detection accuracy, they are able to provide only binary change mapping where are illustrated change regions and non-change regions without any specification of the ground classes nature. In order to get a multiclass change detection leading to a more accurate change mapping, we proposed in this paper a fully unsupervised three steps methodology where during the first step, a binary change mapping is obtained using k-MAD (kernel multivariate alteration detection) components combined with a Chi- 2 test thresholding, during the second step, the change and non-change regions are iteratively classified due to the AP (affinity propagation) clustering algorithm to reach multiclass non-change area and “from-to” change area, finally in the third step, samples from the changed and unchanged classes are involved in a DNN (deep neural network) architecture to provide a multiclass land change mapping. Co-registered bi-temporal multispectral images taken over the northeastern region of Algiers, Algeria, between 1997 and 2001 by the American LANDSAT-TM satellite are used to verify the effectiveness of the proposed scheme. The obtained “from to” land change map validated by means of the spectral signature analysis is more informative on the change and non-change pixels comparing to the binary land change map.