基于Sentinel-2影像的淡水养殖水生动物类型识别研究  

Identification of Farmed Aquatic Animals Types in Freshwater Aquaculture Ponds Based on Sentinel-2 Imagery

作　　者：金晶 毛星 张欣 刘杨[1] 陆学文[1] 任妮[1] JIN Jing;MAO Xing;ZHANG Xin;LIU Yang;LU Xuewen;REN Ni(Information Center,Jiangsu Academy of Agricultural Sciences,Nanjing 210014,China)

机构地区：[1]江苏省农业科学院信息中心,江苏南京210014

出　　处：《河南农业科学》2022年第4期160-170,共11页Journal of Henan Agricultural Sciences

基　　金：江苏省农业科技自主创新资金项目[CX(19)1003]。

摘　　要：为了利用遥感影像实现内陆淡水养殖空间分布的快速提取,以江苏省宜兴市为研究区域,基于Sentinel-2卫星影像数据,提出了一种结合卷积神经网络和随机森林算法的内陆淡水养殖池塘水产类型的识别方法。该方法以深度学习为基础,构建卷积神经网络模型进行养殖池塘语义分割,进而分析养殖区域斑块的归一化植被指数和归一化水体指数,最后采用随机森林算法区分养殖池塘的水产类型。结果表明,宜兴市2021年淡水养殖池塘面积为121.25 km^(2),其中蟹塘面积74.48 km^(2),鱼塘面积46.77 km^(2),识别总体精度为88.33%,kappa系数为0.8243。In order to realize the rapid extraction of the spatial distribution of inland freshwater aquaculture based on remote sensing images,this study chose Yixing City,Jiangsu Province as the study area and proposed a method to identify inland freshwater aquaculture ponds by combining the convolutional neural network and random forest algorithm based on the Sentinel-2 satellite images.Firstly,the study built a convolutional neural network model for semantic segmentation of aquaculture ponds based on deep learning.Then,normalized difference vegetation index(NDVI)and normalized difference water index(NDWI)of the patches in the aquaculture areas were analyzed.Finally,the random forest algorithm was used to distinguish the types of aquaculture ponds.The results showed that there were 121.25 km^(2)of freshwater aquaculture ponds in Yixing in 2021,and the areas of crab ponds and fish ponds were 74.48 km^(2)and 46.77 km^(2),respectively.The overall accuracy of the method was 88.33%,and the kappa coefficient was 0.8243.

关 键 词：淡水养殖池塘 Sentinel-2遥感影像 卷积神经网络 随机森林 SE-Unet 

分 类 号：TP79[自动化与计算机技术—检测技术与自动化装置]