基于U-Net模型的无人机影像数据地表覆被信息自动提取研究

Automatic Extraction of Land Cover Information from UAV Image Data Based on U-Net Model

基于深度学习和无人机获取的高分辨率影像是快速提取林地等地表覆被空间信息的一种新方法。为给后续的林地覆被空间信息自动提取提供技术参考,以南宁市上林县某村的甘蔗(Saccharum officinarum)种植区为研究对象,采用UNet模型对研究区无人机正射影像进行甘蔗种植区的空间位置信息提取。针对目标区域样本数据缺少的问题,首先采用数据增广方式对训练数据集进行增广,然后采用迁移学习的方法在开源数据集预训练U-Net模型,最后利用预训练UNet模型对研究区的影像数据进行训练和测试。结果表明,该方法从正射影像中提取甘蔗种植区空间位置信息的准确率、精确率和召回率分别为98.34%、93.10%和89.21%,不受甘蔗种植区分布密集程度的影响。

A new technique to quickly extract spatial information of land cover,such as forest land,is high-res‐olution image obtained by UAV based on deep learning.In order to provide technical reference for subsequent research on automatic extraction of forest land cover spatial information,taking Saccharum officinarum planting area of a village in Shanglin county,Nanning city as research object,U-Net model was used to extract spatial information of S.officinarum planting area from UAV orthophoto images.Data augmentation was used to aug‐ment training data set,transfer learning was used to pre-trained U-Net model in open source data set,and final‐ly the pre-trained U-Net model was used to train and test image data in study area to solve the problem of lack of sample data in the target area.Results showed that method of extracting spatial information of S.officinarum planting area from orthophotos had an accuracy,a precision rate and a recall rate of 98.34%,93.10%and 89.21%respectively and was not affected by density of S.officinarum planting area.