哈尔滨工程大学学报2025，Vol.46Issue(2)：344-354,11.DOI:10.11990/jheu.202211028
基于自注意力机制的高分遥感影像语义分割
Semantic segmentation of high-resolution remote sensing images using self-attention mechanism
摘要
Abstract
To address the challenges of of multiscale feature extraction and insufficient utilization of context infor-mation in remote sensing images,this paper presents a linear multihead self-attention network model.The model combines a self-attention mechanism with depthwise separable convolution,specifically designed for the semantic segmentation of high-resolution remote sensing images.A depthwise separable convolution module,which reduces computational load and facilitates local feature extraction,is initially introduced.Then,a linear multihead self-at-tention module is proposed in the encoder branch to reduce complexity.Finally,a decoder is designed to restore the resolution of the feature map and integrate multilevel features through a cascade operation,generating high-reso-lution semantic segmentation results.The proposed method achieves mF1 scores of 90.77%and 92.36%on the IS-PRS Vaihingen and Potsdam datasets,respectively.Compared with current mainstream algorithms,the segmenta-tion accuracy and overall accuracy of impervious surfaces,buildings,low vegetation,and trees are improved.The proposed linear multihead self-attention network is an efficient semantic segmentation model for high-resolution re-mote sensing images.关键词
高分辨率遥感影像/多头自注意力/深度可分离卷积/语义分割/特征提取/卷积神经网络/编码器/解码器Key words
high-resolution remote sensing image/multihead self-attention/depthwise separable convolution/se-mantic segmentation/feature extraction/convolutional neural network/encoder/decoder分类
天文与地球科学引用本文复制引用
杨军,张金影,康玥..基于自注意力机制的高分遥感影像语义分割[J].哈尔滨工程大学学报,2025,46(2):344-354,11.基金项目
国家自然科学基金项目(42261067,61862039) （42261067,61862039）
兰州市人才创新创业项目(2020-RC-22) （2020-RC-22）
兰州交通大学天佑创新团队项目(TY202002) （TY202002）
甘肃省教育厅优秀研究生"创新之星"项目(2022CXZX-613). （2022CXZX-613）
