摘要
针对高分辨率遥感影像中彩钢板建筑物具有尺度差异大、类内颜色特征不稳定等问题,提出一种基于改进UNet网络的彩钢板建筑物提取方法(MFEDNet)。在编码器部分,利用新构建的多尺度特征增强模块替代原始UNet双卷积操作,提高模型多尺度特征感知能力的同时能获取更为广泛的上下文信息。在跳跃连接阶段引入双注意力机制增强模型对特征的学习能力,其中的通道注意力模块有效增强了彩钢板建筑物类别颜色信息,解决了类内颜色特征不稳定问题。在自建的GCS数据集和开源CSS数据集上分别进行了对比实验,precision和OA分别达到了87.02%、92.26%和96.55%、92.63%,显著高于对比实验方法。实验结果表明,该方法能够有效提取彩钢板建筑物,避免了误提及空洞现象。
Aiming at the problems of large scale differences and unstable intra-class color features of color steel buildings in high-resolution remote sensing images,this paper proposes a color steel building extraction method based on improved UNet network(MFEDNet).In the encoder part,the newly constructed multi-scale feature enhancement module is used to replace the original UNet double convolution operation,which improves the multi-scale feature perception ability of the model and at the same time obtains a wider range of contextual information.The dual-attention mechanism is introduced in the jump-connection stage to enhance the model’s feature learning ability,in which the channel attention module effectively enhances the color information of the color steel building category and solves the problem of instability of intra-class color features.In this paper,comparative experiments are conducted on the self-built GCS dataset and open source CSS dataset,respectively,and precision and OA reach 87.02%,92.26% and 96.55%,92.63%,respectively,which are significantly higher than that of the comparative experimental methods.The experimental results show that the proposed method is able to extract the color steel buildings effectively,and it avoids misreferring to the void phenomenon.
作者
贺鹏程
杨树文
单文超
杨海燕
HE Pengcheng;YANG Shuwen;SHAN Wenchao;YANG Haiyan(Faculty of Geomatics,Lanzhou Jiaotong University,Lanzhou 730700,China;National and Local Joint Engineering Research Center for the Application of Geographical Monitoring Technology,Lanzhou 730700,China;Gansu Provincial Key Laboratory of Surveying and Mapping Science and Technology,Lanzhou 730700,China)
出处
《遥感信息》
北大核心
2025年第2期87-95,共9页
Remote Sensing Information
基金
国家自然科学基金(42161069)。
关键词
深度学习
高分辨率遥感影像
彩钢板建筑物
提取
条带卷积
双注意力机制
deep learning
high resolution remote sensing imagery
color steel plate building
extraction
strip Convolution
dual attention mechanism
