基于Mamba的双分支遥感图像海陆分割模型

打开文本图片集

中图分类号：TP338.6 文献标志码：A

Abstract： Aiming at the problems low accuracy efficiency deep learning models for segmenting complex coastlines in remote sensing images，an efficient sea-l segmentation model（MRU-Net） for complex types coastlines was proposed. The ResNeXt-50 network MRU-Net as an encoder was employed to extract the rich feature information the image， a multi-scale segmentation attention（MSA） module was designed to enhance the weak features at the sea-l boundary. The SEMamba architecture was introduced as the decoder to capture long-range dependencies integrate high low-level semantic information. Model comparison experiments were conducted based on Benchmark Sea-L Dataset（BSD）， the results visualization experiments showed that the MRU-Net model achieved 97.75% Mean Intersection over Union on BSD，which can achieve more accurate sea-l segmentation results.

Keywords： remote sensing image; sea-l segmentation;multi-scale segmentation attention

遥感图像中包含复杂的场景和纹理分布，空间信息丰富，海岸线两侧的地物具有独特的特征，陆地土壤和近海沉积物的光谱特征也较为相似，传统的海陆分割方法的准确度较低。（剩余11271字）