摘要
针对仅使用无人机遥感RGB影像进行目标检测时精度不高、高程信息利用不足等问题,该文采用通道叠加和类IHS变换两种多通道数据融合方法对RGB影像与高程数据进行融合,使用DeepLabv3+卷积神经网络语义分割模型提取两种融合影像地物目标,并与RGB影像提取结果进行对比分析。结果表明,基于上述两种融合影像的地物目标识别精度高于仅使用RGB影像的识别精度,其中通道叠加影像的整体像素精度、平均像素精度和Kappa系数分别提高了3.52%、1.42%和14.99%。由于不同地物目标与周围地物的高程差不同,致使各融合方法对不同地物目标识别精度的提升效果不同,道路、建筑和地面识别精度的提升效果较好。
To address the problems of low accuracy and insufficient utilization of elevation information in target detection using UAV remote sensing RGB images only,this paper adopts channel superposition and imitates IHS transform to fuse RGB images with elevation data,uses DeepLabv3+convolutional neural network semantic segmentation model to extract the ground targets from the two kinds of fused images,and compares and analyzes the extraction results with that of RGB images.The results show that the recognition accuracy of ground targets based on the above two fused images is higher than that using RGB images alone.The overall pixel accuracy,mean pixel accuracy and Kappa coefficient of channel superimposed images are improved by 3.52%,1.42%and 14.99%,respectively.Due to the different elevation differences between different ground targets and the surroundings,the enhancement effect of each fusion method on the recognition accuracy of various ground targets is different,and the enhancement effect of road,building and other ground recognition accuracy is better.
作者
孙晓宇
蔡祥
SUN Xiao-yu;CAI Xiang(School of Information Science & Technology,Beijing Forestry University,Beijing 100083,China)
出处
《地理与地理信息科学》
CSCD
北大核心
2021年第6期41-45,共5页
Geography and Geo-Information Science
基金
国家重点研发计划项目“西北干旱荒漠区煤炭基地生态安全保障技术”(2017YFC0504400)
“矿区生态修复与生态安全保障技术集成示范研究”(2017YFC0504406)
国家自然科学基金项目“保墒造林技术水分涵养效果见色方法研究”(31400621)。
关键词
数据融合
地物目标提取
深度学习
语义分割
无人机遥感影像
data fusion
ground targets extraction
deep learning
semantic segmentation
UAV remote sensing images
