基于高光谱影像的长三角一体化示范区典型区域植被类型分类  

Vegetation classification in typical areas of the demonstration zone of integrated regional development of Yangtze River Delta based on hyperspectral satellite images

作　　者：潘琛[1] 顾建祥[1] 岳照溪 PAN Chen;GU Jianxiang;YUE Zhaoxi(Shanghai Surveying and Mapping Institute,Shanghai 200063,China)

机构地区：[1]上海市测绘院,上海200063

出　　处：《时空信息学报》2024年第6期698-709,共12页JOURNAL OF SPATIO-TEMPORAL INFORMATION

基　　金：上海市2022年度“科技创新行动计划”高新技术领域项目(22511102800)。

摘　　要：长三角一体化示范区植被具有类型多样、分布格局破碎等特征,给植被空间实地调查带来了一定的困难。本文基于资源一号02D(ZY1E)高光谱卫星影像,建立植被类型光谱库,不仅包含不同季节的植被光谱特征,还充分考虑植被类型的多样性和空间分布的破碎性。基于ShuffleNet轻量级深度学习网络,结合空谱注意力机制,构建高光谱植被分类深度学习网络,利用遥感影像的光谱、空间特征,实现不同季节研究区域植被分类;并与常用方法卷积神经网络(convolution neural network,CNN)、支持向量机(support vector machine,SVM)进行比较分析。结果表明:较常规的CNN和SVM方法,本方法在精确率、召回率、F1和总体精度四项评价指标均有较大提升,能够较好地刻画地物轮廓和保持斑块完整性,且多期分类结果总体精度均能够达到0.85以上;水田作物(以水稻为主)是研究区夏季主要的基本农作物,以茭白为代表的水生植被、林地也有大量分布。Monitoring the spatial pattern of vegetation is a crucial aspect of ecological environment monitoring.However,traditional methods still have significant limitations in large-scale,long-term and dynamic monitoring tasks.Hyperspectral remote sensing technology emerges as a promising solution,addressing these shortcomings while meeting the requirements for meticulous management.The demonstration zone of integrated regional development of Yangtze River Delta prioritizes sustainable ecological green development and places great emphasis on the construction of ecological civilization.Nonetheless,the diversity of vegetation types and fragmented distribution patterns within this demonstration zone pose challenges to investigating vegetation spatial dynamics.Therefore,there is an urgent need for a comprehensive understanding of the ecological environment in this area,particularly in accurately depicting the spatial distribution of vegetation using hyperspectral remote sensing technology.This article introduces a method for vegetation classification and spatial pattern extraction,using ZiYuan102D(ZY1E)hyperspectral satellite images across various seasons.The methodology encompasses hyperspectral satellite image data processing,the establishment of a vegetation spectral library,and the optimization of a deep learning network.Through field investigations of vegetation types at sample points,a comprehensive vegetation spectral library for the demonstration zone is established,capturing the distinct spectral characteristics of vegetation in hyperspectral remote sensing images across different seasons.Utilizing the lightweight deep learning network of ShuffleNet and incorporating the spatial-spectral attention mechanism,a tailored deep learning network for vegetation classification is constructed to realize vegetation classification using hyperspectral satellite images for the studied area.Aiming at the characteristics of diverse vegetation types,spatial distribution and different spectral characteristics,the network incorporat

关 键 词：高光谱影像 植被分类提取 空谱注意力机制 植被光谱库 深度学习 资源一号遥感影像 光谱曲线 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]