Abstract:In traditional visual Transformer hyperspectral unmixing methods, the limitations of fixed-size patch division and the insufficient ability of the model to recognize key information result in the difficulty in effectively capturing multi-scale features and the underutilization of important information in channels. To address the aforementioned issues, a hyperspectral unmixing network scheme based on multi-scale feature fusion and channel calibration is proposed. This scheme introduces a multi-scale feature fusion mechanism, which achieves the fusion of multi-scale features through depthwise separable convolution and dilated convolution, enhancing the ability to capture spatial structural information. Additionally, a dynamic channel recalibration mechanism is introduced to adaptively adjust the weights of channels, strengthen the interaction between different channels, highlight useful information, and suppress redundant noise. Experimental results show that the proposed method exhibits good unmixing performance on multiple real datasets, achieving significant performance improvements compared to traditional methods.