融合多尺度特征的河道型水库网箱养殖信息提取研究  

Research on information extraction of cage aquaculture in river-type reservoir based on multi-scale feature fusion

作　　者：林万强 陈芸芝[1,2,3] LIN Wanqiang;CHEN Yunzhi(Academy of Digital China(Fujian),Fuzhou University,Fuzhou 350108,China;Laboratory of Spatial Data Mining and Information Sharing of Ministry of Education,Fuzhou University,Fuzhou 350108,China;National and Local Joint Engineering Research Center for the Comprehensive Application of Satellite Space Information Technology,Fuzhou 350108,China)

机构地区：[1]福州大学数字中国研究院(福建),福州350108 [2]福州大学空间数据挖掘与信息共享教育部重点实验室,福州350108 [3]卫星空间信息技术综合应用国家地方联合工程研究中心,福州350108

出　　处：《测绘科学》2024年第10期133-145,共13页Science of Surveying and Mapping

基　　金：福建省自然科学基金项目(2022J01111);福建省省属公益类科研院所基本科研专项(2023R1012005)。

摘　　要：为制定适合河道型水库的网箱养殖信息提取方法,实现水库网箱养殖自动化精准提取,该文基于“U”型编解码结构,顾及多尺度特征信息,通过引入残差单元(RU)、高效多尺度注意力(EMA)、改进级联多尺度卷积(MCP)以及嵌入多尺度特征(IAC)等模块改进深度学习网络构建EAMRNet模型,以闽江流域水口库区为研究区,开展水库网箱养殖信息提取研究。结果表明,EAMRNet模型提取的交并比(IoU)、召回率(Recall)、精准率(Precision)、F_(1)分数(F_(1)-score)分别为80.26%、90.94%、87.23%、89.05%,相比于UNet、ResUNet、DeepLab V3+、TransUNet、HRNet等5种经典模型精度评价结果,精度均为最高。同时,将EAMRNet模型应用于提取闽江流域水口库区2019年—2023年网箱养殖信息,经提取结果统计,闽江流域水口库区网箱养殖面积从2019年的333.9652 hm2减少至2023年的156.7713 hm2,总体呈现先增后减的趋势。综上,改进后的模型在水库网箱养殖提取任务上具备较高的提取精度,该研究可以为当地养殖管理部门进行养殖动态监测及合理规划养殖提供理论依据和数据支撑。In order to develop a method for extracting information on cage aquaculture suitable for river-type reservoirs and achieve automated and accurate extraction of cage aquaculture information in reservoirs,this article improves the construction of a deep learning network based on the“U”type encoding and decoding structure,taking into account multi-scale feature information.By introducing modules such as Residual Unit(RU),Efficient Multi-Scale Attention(EMA),improved cascade multi-scale convolution(MCP),and embedded multi-scale features(IAC),to construct the EAMRNet model.Taking the Shuikou reservoir area of Minjiang River Basinn as the research area,the study on extracting information on cage aquaculture in reservoirs was conducted.The results showed that the IoU,Recall,Precision,and F_(1)-score of the EAMRNet model were 80.26%,90.94%,87.23%,and 89.05%,respectively.Compared with the accuracy evaluation results of five classic models such as UNet,ResUNet,DeepLab V3+,TransUNet,and HRNet,the accuracy was the highest.At the same time,the EAMRNet model was applied to extract cage aquaculture information from 2019 to 2023 in the Shuikou reservoir area of Minjiang River basin.According to the statistics of the extraction results,the cage aquaculture area in the Shuikou reservoir area of Minjiang River basin decreased from 333.9652 hm~2 in 2019 to 156.7713 hm~2 in 2023,showing a trend of first increasing and then decreasing.In summary,the improved model has high extraction accuracy for extracting information on cage aquaculture in reservoirs.This study can provide theoretical basis and data support for local aquaculture management departments to conduct dynamic monitoring and rational planning of aquaculture.

关 键 词：河道型水库 多尺度特征 水库网箱养殖 深度学习 国产高分辨率遥感影像 

分 类 号：P237[天文地球—摄影测量与遥感]