Building change detection using the parallel spatial-channel attention block and edge-guided deep network

Abstract

Building change detection in high-resolution satellite images plays a special role in urban management and development. Recently, methods for building change detection have been greatly improved by developing deep learning. Although deep learning technologies, especially Siamese convolutional neural networks, have been successful and popular, they usually have problems in extracting features that are not discriminative enough and also cause the loss of shape and details at the edges. To address these problems, a dual-branch deep network and a parallel spatial-channel attention mechanism were suggested to extract spatial and spectral dependencies and more discriminative features. The spatial attention unit measured the rich context of local features, and the distinction between changed objects and backgrounds was increased using spatial attention in deep features. The channel attention module adjusted the weight of channels and acted as a channel selection process. Mixing two attentions to parallel mode made the features more practical, and useful information was learned more robustly. Moreover, a dual loss function was proposed in which the edge-based consistency constraints were used in the first part to converge the edges of the training and the predicted data. The weighted binary cross-entropy was added to the second part of the loss function. The proposed method was implemented on two remote sensing datasets, and the results were evaluated with state-of-the-art methods. With the proposed model, the F1-score was improved by 2.43% and 1.83% in the first and second datasets, respectively.