结合VGGNet与Mask R-CNN的高分辨率遥感影像建设用地检测  

High Resolution Remote Sensing Image Construction Land Detection Combined with VGGNet and Mask R-CNN

作　　者：陈敏 潘佳威 李江杰 徐璐[2] 刘加敏 韩健 陈奕云[2,4] Chen Min;Pan Jiawei;Li Jiangjie;Xu Lu;Liu Jiamin;Han Jian;Chen Yiyun(Guangzhou Urban Planning Survey&Design Survey Research Institute,Guangzhou 510060,China;School of Resource and Environmental Science,Wuhan University,Wuhan 430079,China;Guangxi Huayao Space Information Technology co.LTD,Nanning 530031,China;State Key Laboratory of Soil and Sustainable Agriculture,Chinese Academy of Sciences,Nanjing 210008,China)

机构地区：[1]广州市城市规划勘测设计研究院,广东广州510060 [2]武汉大学资源与环境科学学院,湖北武汉430079 [3]广西华遥空间信息科技有限公司,广西南宁530031 [4]土壤与农业可持续发展国家重点实验室,江苏南京210008

出　　处：《遥感技术与应用》2021年第2期256-264,共9页Remote Sensing Technology and Application

基　　金：国家重点研发计划“绿色宜居村镇技术创新”重点专项项目子课题“村镇发展潜力因子识别与指标信息快速获取技术”(2018YFD1100801-01)。

摘　　要：针对当前多数深度学习模型只能对高分辨率遥感影像裁剪图片进行土地利用类型判别的问题,结合VGGNet与Mask R-CNN开展了智能化建设用地目标检测研究。在建立研究区4类土地利用类型遥感影像数据集的基础上,对比了VGGNet、ResNet和DenseNet 3种卷积神经网络模型的分类精度,选取分类效果最优的神经网络模型VGGNet与Mask R-CNN实现建设用地目标检测智能化。结果表明:(1)VGGNet、ResNet和DenseNet 3种卷积神经网络模型的分类精度分别为:97.44%、93.75%和95.13%,且VGG16模型迭代次数最少,训练时间相对较少;(2)Mask R-CNN阈值设置对目标检测精度有重要的影响,当阈值设定为0.3时,VGG16结合Mask R-CNN的联合模型对建设用地检测的标定框精度最高。同时联合模型比单一使用Mask R-CNN模型对建设用地检测有更高的准确率,并且表现出了更强的适应性和鲁棒性。To address the problem that most current deep learning models can only discriminate land use types for cropped images of high-resolution remote sensing images,this paper combines VGGNet and Mask R-CNN to carry out a study on intelligent construction land target detection.On the basis of establishing remote sensing image datasets of four types of land use types in the study area,we compare the classification accuracy of three convolutional neural network models,VGGNet,ResNet and DenseNet,and select the neural network model with the best classification effect,VGGNet and Mask R-CNN,to achieve intelligent construction land target detection.The results show that:(1)the classification accuracies of the three convolutional neural network models VGGNet,ResNet and DenseNet are 97.44%,93.75%and 95.13%,respectively,and the VGG16 model has the least number of iterations and relatively less training time;(2)the Mask R-CNN threshold setting has an important influence on the target detection accuracy,when the threshold is set to is 0.3,the joint model of VGG16 combined with Mask R-CNN has the highest calibration frame accuracy for construction land detection.Also the joint model has higher accuracy than the single use of Mask R-CNN model for construction land detection,and shows more adaptability and robustness.

关 键 词：卷积神经网络(CNN) 目标检测 影像分类 高分辨率遥感影像 建设用地 

分 类 号：TP79[自动化与计算机技术—检测技术与自动化装置]