基于U-Net卷积神经网络的广东省海水养殖区识别及其时空变化遥感监测  

Identification of mariculture areas in Guangdong Province and remote sensing monitoring of their spatial and temporal changes based on the U-Net convolutional neural network

作　　者：苏玮 林阳阳 岳文 陈颖彪[2] SU Wei;LIN Yangyang;YUE Wen;CHEN Yingbiao(Land Investigation and Planning Institute of Guangdong Province,Guangzhou 511453,China;School of Geography and Remote Sensing,Guangzhou University,Guangzhou 510006,China)

机构地区：[1]广东省土地调查规划院,广州511453 [2]广州大学地理科学与遥感学院,广州510006

出　　处：《自然资源遥感》2022年第4期33-41,共9页Remote Sensing for Natural Resources

基　　金：广东省海洋综合管理专项项目“广东省养殖用海调查”(编号:440000210000000019287);广东省土地调查规划院立项项目“广东省养殖用海外业调查、成果编制及质量管控/养殖用海补充调查、数据集成管理示范性服务”(编号:GHYFW20210509/GHYFW20210701);教育部人文社科规划基金项目“基于空间博弈理论的粤港澳大湾区生态红线划定规则及情景模拟研究”(编号:21YJAZH009)共同资助。

摘　　要：海水养殖业在广东省海洋经济中占据重要地位,及时准确地掌握海水养殖区的空间分布及面积变化趋势,对海水养殖业的可持续发展具有重要的促进作用。相较于传统解译方法存在可重复性差、适用范围小、主观随意性强等问题,深度学习网络模型中的U-Net卷积神经网络能够更好地从遥感影像中提取目标特征,具有更高的提取精度。鉴于此,基于多时相Landsat TM/OLI遥感影像,选用U-Net模型作为解译模型,识别1998—2021年广东省海水养殖区(围海养殖区及开放式网箱养殖区),开展海水养殖区面积趋势性分析,并探究海水养殖区在空间分布格局上的变化特征。结果表明:相较于K-Means聚类分析和深度信念网络等网络模型,U-Net模型更加适用于对广东省海水养殖区的解译,具有更高的解译精度;广东省海水养殖区主要集中分布在湛江、江门和阳江等广东省西侧区域;广东省各区域海水养殖区面积可分为3个梯队,且变化幅度较小,保持相对稳定状态;广东省海水养殖区在空间上呈现出1998—2014年向外扩张、2014—2021年向内收缩的趋势。本研究能够为广东省海水养殖区的科学管理提供数据支持和技术支撑。The mariculture industry occupies an important position in the marine economy of Guangdong Province.Timely and accurate knowledge of the spatial distribution and area changing trends of mariculture areas can greatly promote the sustainable development of the mariculture industry.Conventional interpretation methods for remote sensing images have problems of poor repeatability,low applicability,and high subjective arbitrariness.By contrast,the U-Net convolutional neural network,which belongs to the deep learning network model,can better extract the features of the object with higher extraction precision.Therefore,based on the multi-temporal Landsat TM/OLI remote sensing images,this study identified the mariculture areas(enclosed-sea and open-cage aquaculture areas)in Guangdong from 1998 to 2021 using the U-Net model as the interpretation model.The area trend analysis of mariculture areas was made.The changing characteristics of mariculture areas in terms of spatial distribution patterns were studied.The results are as follows.Compared with network models such as K-Means cluster analysis and DBN,the U-Net model with higher interpretation precision is more suitable for the interpretation of mariculture areas in Guangdong.The mariculture areas in Guangdong are mainly distributed in the western portion of Guangdong,such as Zhanjiang,Jiangmen,and Yangjiang.The mariculture areas in Guangdong can be classified into three levels in terms of area.They have small changes and keep a relatively stable state.The mariculture areas in Guangdong showed a spatial trend of outward expansion from 1998 to 2014 and inward contraction from 2014 to 2021.This study will provide data and technical support for the scientific management of the mariculture areas in Guangdong.

关 键 词：海水养殖区 遥感 深度学习 U-Net模型 广东省 

分 类 号：TP79[自动化与计算机技术—检测技术与自动化装置]