基于ACSBL-DeepLabV3+的遥感图像地物分类方法研究  

Research on Remote Sensing Image Land Cover Classification Based on ACSBL-DeepLabV3+

作　　者：冯丹亭 于淼[1] 于晓鹏[1] FENG Danting;YU Miao;YU Xiaopeng(College of Mathematics and Computer,Jilin Normal University,Siping 136099,China)

机构地区：[1]吉林师范大学数学与计算机学院,吉林四平136099

出　　处：《无线电工程》2025年第3期548-557,共10页Radio Engineering

基　　金：吉林省科技发展计划项目(YDZJ202301ZYTS285)。

摘　　要：针对高分辨率遥感图像分割中复杂环境下的遥感影像提取不精准及小型物体易被忽略,导致类别提取不完整、物体边界模糊的问题,提出一种基于DeepLabV3+网络改进的遥感图像语义分割方法。在编码器部分,采用MobileNetV2轻量级网络作为主干特征提取网络,使用非对称卷积空间金字塔池化模块(Asymmetric Spatial Pyramid Pooling Module,ACS-ASPP),将解码器进行细化,与主干网络提取的浅层特征加权融合,引入选择性大核注意力(Large Selective Kernel Attention,LSK)机制。在Vaihingen和Potsdam数据集高分辨率遥感影像数据集上的实验表明,该方法多项性能评价指标均优于U-Net、PSP-Net、全卷积网络(Fully Con-volutional Network,FCN)等多个语义分割网络,总体平均交并比(mean Intersection over Union,mIoU)分别达到69.13%、75.68%,F1-Score分别达到80.75%、85.84%。实验结果表明,该网络能够有效对各个类别进行分类,具备较高的实用价值。In high-resolution remote sensing image segmentation,DeepLabV3+network has many problems such as large parameters,high training cost,and inaccurate extraction of remote sensing images in complex environments,and small objects are easy to be neglected,which leads to incomplete category extraction and fuzzy boundaries.To address the above problems,an enhanced method based on DeepLabV3+model is proposed for remote sensing image semantic segmentation.In the encoder,the lightweight network MobileNetV2 is used as the backbone feature extraction network,and an Asymmetric Spatial Pyramid Pooling Module(ACS-ASPP)is incorporated.In decoder refinement,the shallow features extracted by the backbone network are continuously added for feature fusion,and Large Selective Kernel Attention(LSK)mechanism is introduced.Experiments on two high-resolution remote sensing image datasets,Vaihingen and Potsdam,demonstrate that the proposed method is superior to several semantic segmentation networks such as U-Net,PSP-Net and Fully Convolutional Network(FCN)in several performance evaluation indexes.The overall mIoU reaches 69.13%and 75.68%,and F1-Score reaches 80.75%and 85.84%,respectively.The experimental results show that the network can effectively classify various categories of objects and has high practical value.

关 键 词：遥感图像多分类 语义分割 DeepLabV3+ 选择性大核注意力机制 解码器细化 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]