一种无人机正射影像地类自动识别方法  

A Method for Automatic Land Classification of UAV Digital Orthophoto Map

作　　者：金振阳 章迪[2] 方田野 石淼 JIN Zhenyang;ZHANG Di;FANG Tianye;SHI Miao(Guangzhou Quancheng Multidimensional Information Technology Company,Guangzhou 511457,China;School of Geodesy and Geomatics,Wuhan 430079,China;Wuhan Geomatics Institute,Wuhan 430022,China)

机构地区：[1]广州全成多维信息技术有限公司,广东广州511457 [2]武汉大学测绘学院,湖北武汉430079 [3]武汉市测绘研究院,湖北武汉430022

出　　处：《城市勘测》2024年第3期83-87,共5页Urban Geotechnical Investigation & Surveying

基　　金：湖北省自然科学基金计划项目资助(2022CFB090);中央高校基本科研业务费专项资金资助(2042023kf0002);第三次全国土地调查项目(445323-201809-324002-0019)。

摘　　要：当前,对无人机正射影像的地类识别仍主要依靠人机交互的方式进行,受作业人员熟练度制约,生产效率较低、生产成本较高。为此,提出了一种无人机正射影像地类自动识别方法,即利用VGG Image Annotator对影像地类按最小单元标注获得精细化样本;搭建以ResNet50为特征提取器的Mask R-CNN网络;基于预训练模型、利用地类样本对网络进行训练和测试。利用某地1 m分辨率的无人机数字正射影像制作了房屋、耕地、森林、水域四种地类样本,依托TensorFlow-gpu 1.11.0和Keras2.0.9搭建训练和测试环境,结果表明,四种地类识别的F1值可达70%以上,证明了本方法的可行性。The land classification of UAV digital orthophoto map still mainly depends on human-computer interaction,which requires experienced and qualified operators,with low efficiency and high cost.Therefore,a method for automatic land classification of UAV Digital Orthophoto Map was proposed,using VGG Image Annotator to divide the land types in the image into minimum units to obtain refined samples;building a Mask R-CNN network with ResNet 50 as the feature extractor;training and testing the network using pre-trained models and land type samples.Sample set was produced using UAV digital orthophoto image of a certain region,with 1m resolution,and a training and testing environment was built based on TensorFlowgpu 1.11.0 and Keras 2.0.9.The test results show that the F1 value of this method for identifying four land types,including house,plowland,forest and water,can reach over 70%,which demonstrates that the new method was feasible.

关 键 词：无人机 数字正射影像 地类识别 Mask R-CNN 

分 类 号：TP753[自动化与计算机技术—检测技术与自动化装置]