基于改进U-Net的典型地物要素提取  

Extraction of Typical Features Based on Improved U-Net

作　　者：马随阳 余永周 望曹俊杰 MA Suiyang;YU Yongzhou;WANG Caojunjie(Changjiang Yichang Waterway Bureau,Yichang 443000,China)

机构地区：[1]长江宜昌航道局,湖北宜昌443000

出　　处：《测绘与空间地理信息》2023年第5期82-85,共4页Geomatics & Spatial Information Technology

摘　　要：针对高分辨遥感影像在进行地物要素提取时,自身带来的影像信息细节化程度高,导致不同地物光谱互相重叠,同种地物光谱分布也是可变的,使得不同地物的相对可分性降低这一问题,本文以高分二号卫星为数据源,提出改进U-Net模型,通过加深U-Net网络结构,引入SFAM模块和ASPP模块,多级尺度特征聚合金字塔方法等对原始U-Net模型进行改进。实验结果显示:改进U-Net模型的总体分类精度OA为88.76%,均交并比MI-oU为0.53,相比原始U-Net模型,FCN模型和SegNet模型的分类精度都有明显的提升。综上可知,本文提出的改进U-Net模型在地物要素提取中是可行的,可为地物要素的高精度提取提供技术支持。Aiming at the problem that the high-resolution remote sensing images bring high degree of detail of image information when extracting features,which leads to overlapping spectra of different features and variable spectral distribution of the same feature,which reduces the relative separability of different features.In this paper,taking Gaofen-2 satellite images as the data source,an improved U-Net model is proposed.The original U-Net model is improved by deepening the structure of U-Net network,introducing SFAM module and ASPP module,and multi-scale feature aggregation pyramid method.The experimental results show that the overall classification accuracy(OA)of the improved U-Net model is 88.76%,and the average intersection ratio MIoU is 0.53,which is significantly improved compared with the original U-Net model,FCN model and SegNet model.To sum up,the improved U-Net model proposed in this paper is feasible in feature extraction,and can provide technical support for high-accuracy feature extraction.

关 键 词：深度学习 语义分割 地物要素 高分二号 

分 类 号：P237[天文地球—摄影测量与遥感]