| 基于Deeplabv3+模型的成都平原水产养殖水体信息提取 |
| |
| 引用本文： | 苟杰松,蒋怡,李宗南,董秀春,吴柏清,刘忠友.基于Deeplabv3+模型的成都平原水产养殖水体信息提取[J].中国农机化学报,2021(3). |
| |
| 作者姓名： | 苟杰松 蒋怡 李宗南 董秀春 吴柏清 刘忠友 |
| |
| 作者单位： | 四川省农业科学院遥感应用研究所;成都理工大学旅游与城乡规划学院 |
| |
| 基金项目： | 四川省科技厅应用基础研究项目(2019YJ0608);四川省农业科学院前沿学科研究基金(2019QYXK036)。 |
| |
| 摘 要： | 为应用深度学习和遥感影像实现养殖水体信息的快速提取,以成都平原为研究区,以Sentinel 2A和高分6号多光谱影像为数据源,基于国产开源深度学习平台PaddlePaddle训练Deeplabv3+语义分割模型,构建遥感影像的水体语义分割模型,用于提取成都平原养殖水体信息。Deeplabv3+方法的总体精度和Kappa系数分别达到94.14%和0.88,均高于归一化差分水体指数法和最大似然监督分类法;模型对阴影和建筑物等误分为水体的抑制效果较好,而对小面积和细小线状水体信息的提取则受影像分辨率影响,效果无明显改进;成都平原2018年和2020年养殖水体面积分别为22.3 khm2和28.6 khm2,其验证区青白江区、新津县和广汉市养殖水体面积的泛化提取结果验证误差均≤±10%。该研究结果可为应用深度学习平台建立遥感影像的水体语义分割模型及提取水产养殖水体信息提供参考。
|
| 关 键 词： | 遥感 深度学习 水体信息 水产养殖 成都平原 |
Aquaculture water body information extraction in the Chengdu plain based on Deeplabv3+model |
| |
| Gou Jiesong,Jiang Yi,Li Zongnan,Dong Xiuchun,Wu Baiqing,Liu Zhongyou.Aquaculture water body information extraction in the Chengdu plain based on Deeplabv3+model[J].Chinese Agricultural Mechanization,2021(3). |
| |
| Authors: | Gou Jiesong Jiang Yi Li Zongnan Dong Xiuchun Wu Baiqing Liu Zhongyou |
| |
| Institution: | (Institute of Remote Sensing Application,Sichuan Academy of Agricultural Sciences,Chengdu,610066,China;College of Tourism and Urban and Rural Planning,Chengdu University of Technology,Chengdu,610059,China) |
| |
| Abstract: | Improving the ability to apply deep learning and remote sensing images to quickly extract aquaculture water body information will play a significant role in scientific aquaculture management and fishery information.The Chengdu plain was selected as research area and Sentinel 2A and GF-6 multi-spectral images as data sources.The Deeplabv3+model was trained based on the domestic open source deep learning platform-PaddlePaddle and a water body semantic segmentation model of remote sensing images was constructed to extract aquaculture water body information in the Chengdu plain.The confusion matrix of three classification method,including two comparative experiments(Normalized Difference Water Index and Maximum Likelihood)was obtained by calculating the difference between classification results and the ground truths resampled from 0.5 m high-resolution water classification results to 10 m as the verification standard.The overall accuracy and Kappa coefficient of Deeplabv3+model reached 94.14%and 0.88 respectively,which performed better than others in extracting water body information from Sentinel 2A.In addition,Deeplabv3+also had obvious elimination effect on the misclassification of water bodies such as shadows and buildings.However,the Deeplabv3+method also had a shortcoming,not good for the extraction of small area and small linear water bodies due to the lower image resolution.The area of aquaculture water in the Chengdu Plain in 2018 and 2020 was 22.3 khm2 and 28.6 khm2 respectively.In verification areas,including Qingbaijiang District,Xinjin County and Guanghan City,the verification errors of the generalized extraction results of the aquaculture water areas were all≤±10%.The research results provide a reference for related research on the application of deep learning platforms to establish water semantic segmentation model based on remote sensing images and extract aquaculture water information. |
| |
| Keywords: | remote sensing deep learning water information aquaculture Chengdu plain |
| 本文献已被 CNKI 维普 等数据库收录！ |
|
