Abstract
Change detection (CD) is the process of acquiring changes in the ground surface by analyzing remotely sensed images of the same location at two different stages. Deep learning is becoming increasingly popular for change detection tasks in remote sensing due to its significant advantages in deep feature representation and nonlinear problem modeling. However, many previous neural network-based approaches require a large number of parameters and computations and high-performance hardware, which makes their practical application in remote sensing challenging. Some lightweight change detection techniques are implemented by enhancing deep features and then extracting their difference information, which tends to ignore many shallow features. We propose a lightweight network MobileNetV2 as an encoder and a modified UNet as a decoder, where MobileNetV2 extracts features from bit-time images and UNet performs layer-by-layer fusion of different stages of difference images to enhance the representativeness of changes. Since MobileNetV2 is used as the encoder of UNet, the computation of the proposed method is greatly reduced. Experiments show that our method has the lowest computational cost (2.38G) and the fewest parameters (2.95M) compared to the current mainstream lightweight networks (FC-Siam-diff32, FC-Siam-conc, A2Net and BIT). Three remote sensing public datasets, SYSU-CD, BCDD and LEVIR-CD, are introduced to validate the proposed method in this paper, and the results show that the F1 of the method in this paper reaches 82.84%, 94.51% and 90.89% respectively. All code as well as detailed instructions are available at https://github.com/tawneydaylily/Mobile-CDNet .