A Lightweight Patch-Level Change Detection Network Based on Multilayer Feature Compression and Sensitivity-Guided Network Pruning

Existing satellite remote sensing change detection (CD) methods often crop large-scale bitemporal image pairs into small patch pairs and then use pixel-level CD methods for fair processing. However, due to the sparsity of change, existing pixel-level methods suffer from a waste of computational cost and memory resources on many unchanged areas, which reduces the processing efficiency on hardware platforms with extremely limited computation and memory resources. To address this issue, we propose a lightweight patch-level CD network (LPCDNet) to rapidly remove a lot of unchanged patch pairs in large-scale bitemporal optical image pairs, helping to accelerate the subsequent pixel-level CD process and reduce memory cost. In our LPCDNet, a sensitivity-guided network pruning method is proposed to remove unimportant channels and construct the lightweight backbone network based on the ResNet18 network. Then, the multilayer feature compression (MLFC) module with multiscale max-pooling structure is designed to compress and fuse the multilevel feature information of image patches. The output of MLFC module is fed into the fully connected decision network to generate the predicted binary label. Finally, a weighted cross-entropy loss is utilized in the training process to tackle the change/unchanged class imbalance problem. Experiments on two CD datasets demonstrate that our LPCDNet achieves more than 1000 frames/s on an edge computation platform, i.e., NVIDIA Jetson AGX Orin, which is more than three times that of the existing methods without noticeable performance loss. In addition, the computational cost of the pixel-level CD processing stage can be reduced by more than 60%.