Abstract:

[Objective] Debris flow disasters, due to their suddenness and extreme destructiveness, have become a key focus in emergency rescue. Low-altitude unmanned aerial vehicle (UAV) remote sensing can efficiently obtain high-resolution image data of disaster-stricken areas. However, how to perform accurate and efficient semantic segmentation of disaster-affected areas remains a common technical challenge in research. [Methods] This paper proposes a lightweight frequency-guided remote sensing segmentation network. [Results] The results show that: (1) Under the premise of maintaining low computational cost, the proposed method efficiently extracts the complex boundary structures of objects in high-resolution images, and the overall performance of the model is improved by approximately 1.98% and 4.43% compared to U-Net and UKAN methods; (2) The Fast Fourier Transform (FFT) maps spatial features to the frequency domain, and by jointly modeling the real and imaginary parts, it effectively captures long-range dependencies and periodic geometric information, compensating for the limitations of traditional convolution in global modeling; (3) The introduction of global semantic tokens establishes a gated fusion mechanism to adaptively regulate the weights between local features and global priors, effectively alleviating the problem of detail loss during the upsampling process. [Conclusions] This research can provide effective technical support for rapid damage assessment of debris flow disasters.

Key words: semantic segmentation, deep learning, frequency-guided features, debris flow disaster