Convolutional Autoencoder-Based Image Reconstruction for Unsupervised Multimodal Change Detection

Abstract

Due to the numerous technological developments, the last years have witnessed an increase in the diversity of remote sensing data, whereas the need to interpret multimodal remote sensing data emerged. In this article, we propose a new multimodal unsupervised change detection strategy that projects a pre-event image in the post-event imaging modality. From a reconstruction perspective, the blocks in the pre-event scene are rebuilt from denoised versions of post-event blocks by means of convolutional denoising autoencoders. In order to perform this reconstruction, a dictionary of locations of similar blocks is learned from the pre-event image by analyzing compressed representations. The experiments, conducted over remote sensing images acquired by different sensors, show the effectiveness and the reliability of the proposed approach in various scenarios reflecting diverse types of changes.