Remote Sensing Image Change Detection Based on Similarity Sensing Network

Remote sensing image change detection has shown its potential in various applications such as ecosystem detection and climate emergency response. However, existing change detection methods fail to adequately address the gradient disappearance caused by class imbalance and ignore the similarity between the dual-temporal images, thereby limiting their learning ability. Aiming at these issues, this paper introduces the ideas of metric learning and attention mechanism into the deep learning network, and proposes a remote sensing image change detection model with two novel designs. First, a similarity-guided attention mechanism is introduced into each stage of the encoder and decoder to adjust the feature weights of different regions, which can better capture the subtle changes of ground cover. Second, the batch-balanced contrastive loss is used to optimize the learning effect of the model on rare categories and difficult samples. In the experiments, the proposed model shows high learning performance and anti-interference ability, which provides an efficient tool for handling change detection tasks with remote sensing images.