摘要
为对农用地(耕地)遥感影像中道路和农田信息进行精确高效的提取,采用卷积神经网络(CNN)的方法,以河北省献县某乡冬小麦种植田为研究区,建立“道路-背景”和“农田-背景”2个高精度遥感影像数据集,构建基于MobileNet v1的U-Net、SegNet、PSPNet、DeepLab v3+和基于MobileNet v2的DeepLab v3+共5种CNN语义分割模型,进行道路和农田提取试验;在模型训练前后加入迁移学习、图像拼接和模型融合3种策略。结果表明:1)在2个数据集上,基于MobileNet v1的U-Net和基于MobileNet v1的SegNet 2种模型的识别率和稳定性最佳;2)在提取道路和农田时,融合后模型的平均交并比值分别为0.8533和0.9568;3)对预测图进行后处理,可以为路径规划和作物秸秆产量计算等研究提供道路拓扑图和农田预测图。
In order to accurately and efficiently extract road and field information from remote sensing image of agricultural land(cultivated land),the Convolutional Neural Network(CNN)method was used,and a winter wheat planting field in a township in Xian County,Hebei Province was taken as the research area to establish two highprecision remote sensing image data sets:“Road-background”and“Field-Background”,Five CNN semantic segmentation models are established:U-Net/SegNet/PSPNet/DeepLab v3+based on Mobilenet v1and DeepLab v3+based on MobileNet v2,Road and field extraction experiments are carried out;Three strategies of transfer learning,image mosaic and model ensemble are added before and after the model training.The results show that:1)The U-Net/SegNet based on MobileNet v1has the best precision and stability on the two data sets.2)When extracting roads and field,the mIoU values of the fused model can reach 0.8533and 0.9568,respectively.3)Post-processing the forecast map can provide road topology maps and field forecast maps for research on path planning and crop straw yield calculations.
作者
陈理
杨广
刘名洋
周宇光
CHEN Li;YANG Guang;LIU Mingyang;ZHOU Yuguang(College of Engineering,China Agricultural University,Beijing 100083,China)
出处
《中国农业大学学报》
CAS
CSCD
北大核心
2022年第6期182-191,共10页
Journal of China Agricultural University
基金
科技部创新方法工作专项项目(2020IM020901)。
关键词
道路和农田提取
遥感影像
深度学习
卷积神经网络
语义分割
图像处理
road and field extraction
remote sensing image
deep learning
convolutional neural network
semantic segmentation
image processing
