基于多尺度特征融合的遥感图像水体分割

李益

基于多尺度特征融合的遥感图像水体分割

引用本文：李益.基于多尺度特征融合的遥感图像水体分割[J].计算技术与自动化,2025,(3):88-93

摘要点击次数: 458

全文下载次数: 226

作者	单位
李益	（江苏科技大学,江苏镇江 212000)

中文摘要:水体语义分割在遥感和地球观测领域发挥着至关重要的作用，现有方法难以在复杂情景下有效分割多尺度水体的细小边界，并且面临标注数据成本高的问题。为此，提出了一种基于多尺度特征融合的水体遥感图像语义分割算法。基于双注意力模块和转换注意力模块设计了水体分割网络主体架构，并设计一种多尺度特征融合模块减少编码器和解码器中不同尺度特征之间的语义差距。在模型训练之前，首先通过基于掩码-重建的自监督学习对编码器进行预训练，以获得具有丰富特征表示的初始化权重。在FloodNet数据集上所提出的方法达到了最高的89.77%mIoU，相比于目前最先进的语义分割算法DAE-Former实现了1.31%的提升。实验结果表明，所提出方法具有较高的准确性。

中文关键词:多尺度特征融合注意力语义分割遥感图像

Remote Sensing Image Water Body Segmentation Based on Multi-scale Feature Fusion

Abstract:Water area semantic segmentation plays a crucial role in remote sensing and earth observation. Existing methods can hardly effectively segment small boundaries of multi-scale water areas in complex scenarios, and face the challenge of high annotation data costs. Consequently, a multi-scale fusion based water area segmentation method for remote sensing images is proposed. The proposed water segmentation network is designed based on dual attention modules and transformed attention modules. A multi-scale feature fusion module is introduced to reduce the semantic gap between different scale features in the encoder and decoder. Prior to model training, the encoder is pre-trained with mask-reconstruction based self-supervised learning to obtain initial weights with rich feature representations. Experimental evaluation was conducted on the FloodNet dataset. The proposed method achieved the highest mIoU of 89.77%, which is a 1.31% improvement compared to the state-of-the-art semantic segmentation algorithm DAE-Former. The experimental results demonstrate the accuracy of the proposed method.

keywords:multi-scale feature fusion attention semantic segmentation remote sensing images

查看全文 查看/发表评论 下载pdf阅读器