基于IA_UNet模型的陇西黄土高原土地覆被分类  

Land cover classification of Longxi Loess Plateau based on the IA_UNet model

作　　者：胡丛慧 刘勇[1] 侯建西 刘东杰 刘怡 HU Cong-hui;LIU Yong;HOU Jian-xi;LIU Dong-jie;LIU Yi(College of Earth and Environmental Sciences,Lanzhou University,Lanzhou 730000,China;Hebei Changfeng Information Technology Co.Ltd.,Shijiazhuang 050000,China)

机构地区：[1]兰州大学资源环境学院,兰州730000 [2]河北长风信息技术有限公司,石家庄050000

出　　处：《兰州大学学报（自然科学版）》2023年第3期295-302,共8页Journal of Lanzhou University(Natural Sciences)

基　　金：国家自然科学基金项目(41271360)。

摘　　要：将Landsat 8 OLI中的2~7波段与缨帽变换得到的亮度、绿度、湿度分量结合作为输入数据组合,在注意力U-Net模型基础上引入位置注意力模块与通道注意力模块,得到改进的注意力U-Net(IA_UNet)模型.结果表明, IA_UNet可以有效地优化U-Net语义分割模型的土地覆被分类效能,加权交并比和总精度比注意力U-Net、 FCN、 ResNet-FCN、 U-Net、 Deeplab V3+模型分别提高了0.44%~3.23%和0.25%~2.09%;缨帽变换特征分量的引入有利于分类精度的提高,其分类结果中加权交并比、总精度比仅使用Landsat 8 OLI中2~7波段分别提高了1.17%、 0.73%,比缨帽变换三分量组合分别提高了1.92%、 1.31%.The 2-7 bands in Landsat 8 OLI were combined with the brightness,greenness and humidity components obtained from tassel hat transformation as the input data combination.Based on the attention U-Net model,the position attention module and channel attention module were introduced to obtain an improved attention U-Net(IA_UNet)model.The results showed that IA_UNet could effectively optimize the land cover classification efficiency of the U-Net semantic segmentation model.The frequency weight-ed intersection over union and total accuracy were 0.44%-3.23%and 0.25%-2.09%higher than attention U-Net,FCN,ResNet-FCN,U-Net and Deeplab V3+models respectively.The introduction of the tassel transform feature component was conducive to the improvement of the classification accuracy.In the clas-sified results,the frequency weighted intersection over union and total accuracy were 1.17%and 0.73%higher than those in band 2-7 of Landsat 8 OLI alone,and 1.92%and 1.31%higher than a combination of the three components of the tassel transform.

关 键 词：土地覆被 缨帽变换 注意力机制 陇西黄土高原 Landsat 8 OLI 

分 类 号：P237[天文地球—摄影测量与遥感]