Abstract:
Objectives The state space model Mamba, which combines global sensing field and dynamic weighting, can better capture the trend and periodic information in the time series prediction task, and has become an important direction in the current deep learning technologies for remote sensing image analysis and interpretation. However, the current research of Mamba in the dense prediction of remote sensing images is insufficient, and there are problems such as high computational complexity and low detection efficiency for the change detection of high-resolution remote sensing images.
Methods We conduct an in-depth analysis of the key factors affecting the number of parameters of the Mamba layer, and propose a dual-stream parallel omnidirectional scan Mamba（DSPOSM）model architecture to construct an in-channel data parallelization processing method. The proposed DSPOSM model can effectively reduce the number of channels of the data in single Mamba block while the total number of channels remains unchanged, in order to achieve highly efficient feature extraction based on the decrease of the overall number of parameters of the model.
Results Comparative experimental results on the LEVIR-CD and WHU-CD datasets show that the proposed method outperforms the non-Mamba architecture method in all metrics and significantly reduces the number of parameters and computational complexity of the network by 35.8% and 22.4%, and improves the training efficiency by 19.45% and 8.26%, compared with the benchmark method, respectively.
Conclusions The proposed DSPOSM method can significantly reduce the number of parameters and computational complexity of Mamba-based networks and improve training efficiency.
下载:
