A lightweight building change detection network with coordinate attention and multiscale fusion

Real-time and rapid detection of building changes can play an important role in natural resource management. In practical change detection, many existing change detection methods need to be further improved, for example, in terms of reducing the computational cost of modeling and improving model accuracy. In this article, we propose a lightweight Siamese multiscale fusion network for building change detection. The encoder uses coordinate attention mechanism enhancement feature extraction, while the decoder uses multi-scale fusion to reconstruct the change region. In the network structure, 3 × 1, 1 × 3 and 1 × 1 convolution is widely used to replace 3 × 3 to reduce network model parameters. Finally, we use a hybrid loss function to alleviate the sample imbalance problem and a deep supervision strategy to fully utilize network hidden layer information. In comparison to other models, the model proposed in this paper achieves the best scores of 90.25