摘要
针对目前的遥感影像分类方法提取整个城市建成区绿地耗时耗力,提取结果容易出现误判的问题,该文提出了一种轻量级PSPNet网络的城市建成区绿地智能提取方法。以MobileNetV2作为特征提取主干网络,并引入深度可分离卷积替代标准卷积,以提升网络的运算速度。利用金字塔池化模块内不同尺度的池化核对MobileNetV2提取的底层特征进行下采样,以获取多尺度特征,然后通过上采样及高层特征融合,形成具有上下文信息的最终特征,从而提高网络的识别精度。在无人机影像上进行实验,并将结果与3种机器学习方法以及5种深度学习网络作对比。结果表明:该文方法优于文中的其他方法,其总体精度达到93.67%,模型训练时间最短,且具备一定的迁移能力和实用性,能快速、精确的提取整个城市建成区绿地。
For the problem of time-consuming and labor-intensive extraction of green space in the entire urban built-up area,as well as the misclassification results obtained by current remote sensing image classification methods,this paper proposes an intelligent extraction method for green space in urban built-up areas based on a lightweight PSPNet network.MobileNetV2 is employed as the backbone network for feature extraction,and depth-wise separable convolutions are introduced to replace standard convolutions,thereby enhancing the computational speed of the network.Different scales of pooling kernels within the pyramid pooling module are applied to downsample the bottom-layer features extracted by MobileNetV2.This process enables the acquisition of multi-scale features.Subsequently,upsampling and high-level feature fusion are performed to generate final features that encompass contextual information.This approach aims to enhance the network's recognition accuracy.The experiments are conducted on UAV images,and the results are compared with three machine learning methods(ML,SVM,RF)and five deep learning networks(VGGNet,U-Net,DeepLabV3+,ShuffleNetV2,PSPNet).The results show that the method of this paper is better than the other methods in the paper,its overall accuracy reaches 93.67%,the model training time is the shortest,and it has a certain degree of migration ability and practicality,and it can quickly and accurately extract the green space of the whole urban built-up area.
作者
曹乾洋
杨广斌
王仁儒
李蔓
骆耀培
陶倩
CAO Qianyang;YANG Guangbin;WANG Renru;LI Man;LUO Yaopei;TAO Qian(School of Geography and Environmental Science,Guizhou Normal University,Guiyang 550025,China)
出处
《测绘科学》
CSCD
北大核心
2023年第9期99-109,共11页
Science of Surveying and Mapping
基金
贵州省科技计划项目
黔科合重大专项[2022]001。
关键词
绿地提取
深度学习
语义分割
PSPNet网络
greenfield extraction
deep learning
semantic segmentation
PSPNet network