The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences
Publications Copernicus
Articles | Volume XLVIII-4/W2-2022
12 Jan 2023
 | 12 Jan 2023


A. Eftekhari, F. Samadzadegan, and F. Dadrass Javan

Keywords: Remote Sensing Image Change Detection, Deep Learning, Attention Mechanism, W-shape Networks, High-resolution Images, Dual Loss Function

Abstract. Building change detection in high resolution remote sensing images is one of the most important and applied topics in urban management and urban planning. Different environmental illumination conditions and registration problem are the most error resource in the bitemporal images that will cause pseudochanges in results. On the other hand, the use of deep learning technologies especially convolutional neural networks (CNNs) has been successful and considered, but usually causes the loss of shape and detail at the edges. Accordingly, we propose a W-shape ResUnet++ network in which images with different environmental conditions enter the network independently. ResUnet++ is a network with residual blocks, triple attention blocks and Atrous Spatial Pyramidal Pooling. ResUnet++ is used on both sides of the network to extract deeper and discriminator features. This improves the channel and spatial inter-dependencies, while at the same time reducing the computational cost. After that, the Euclidean distance between the features is computed and the deconvolution is done. Also, a dual loss function is designed that used the weighted binary cross entropy to solve the unbalance between the changed and unchanged data in change detection training data and in the second part, we used the mask–boundary consistency constraints that the condition of converging the edges of the training data and the predicted edge in the loss function has been added. We implemented the proposed method on two remote sensing datasets and then compared the results with state-of-the-art methods. The F1 score improved 1.52 % and 4.22 % by using the proposed model in the first and second dataset, respectively.