Multiscale Hierarchical Losses to Preserve Hidden-Layer Features for Unsupervised Change Detection

Deep learning (DL) approaches are widely used to improve change detection (CD). In many application domains, unsupervised DL CD methods are preferred since gathering multitemporal labeled samples is challenging. Many unsupervised CD methods use pre-trained DL models to extract multiscale features. This does not allow for preserving the spatial context information and the object structure in the multiscale hidden-layer features, thus obtaining poor performance in modeling multiresolution changes. In this article, we propose two hierarchical loss functions to train multiscale hidden-layer features and preserve their spatial context information. The multiscale hidden-layer feature maps extracted from the model are used in an unsupervised multiscale CD method. We present two possible hierarchical loss functions. The first one considers all the model layers during the training by comparing the mirrored couples of encoder–decoder hidden-layer features, while the second one aims to preserve the geometrical details using a multiresolution-based loss function. After training, the CD method uses a feature selection (FS) based on structure-similarity-index (SSIM) to keep only the most informative hidden-layer feature maps. We tested the proposed method on bi-temporal multispectral images acquired by Landsat-8 representing a burned area and Sentinel-2 images representing a deforested area.