Change Detection Using Siamese Neural Network with Depthwise Separable Convolutions and Spatial Attention Module (SAM) in Remote Sensing.

Remote Sensing is invaluable for land cover monitoring, agricultural surveys, and urban development, particularly in change detection tasks. This research discusses a new approach to change detection utilizing Siamese neural networks integrated with a lightweight architecture comprising Siamese U-Net, residual neural network, depth-wise separable convolution, and spatial attention modules. The aim is to establish a model that achieves high performance without excessive computational parameters. To measure the model, the LEVIR-CD dataset is used, which comprises pairs of images captured at different times. The data in two linked images indicates changes, such as buildings that increase or decrease in a certain area between the two images. We evaluated the performance differences between the proposed model, U-Net, and Siamese U-Net using standard convolution.