基于无人机高光谱影像的城镇不透水面提取  

Extraction of Impervious Surfaces in Towns Based on UAV Hyperspectral Imagery

作　　者：张怡婷 陆冬华 武鼎[1,2] 高岩 ZHANG Yi-ting;LU Dong-hua;WU Ding;GAO Yan(National Key Laboratory of Uranium Resource Exploration-Mining and Nuclear Remote Sensing,Beijing 100029,China;Beijing Research Institute of Uranium Geology,Beijing 100029,China)

机构地区：[1]铀资源探采与核遥感全国重点实验室,北京100029 [2]核工业北京地质研究院,北京100029

出　　处：《光谱学与光谱分析》2025年第4期1150-1158,共9页Spectroscopy and Spectral Analysis

基　　金：国家自然科学基金项目(42202331);中核集团研发平台稳定支持科研项目(WDZC-2023-HDYY-101)资助。

摘　　要：无人机搭载高光谱传感器获取的遥感影像具有光谱信息丰富、空间分辨率高的优势,可为城镇不透水面提取提供更有效的数据。然而,高光谱影像包含大量波段,存在信息冗余,会增加模型训练的复杂度,随着数据维度增加,数据空间体积呈指数级增长,有限样本量在高维空间中会稀疏分布,易导致模型过拟合。此外,传统提取方法特征学习能力有限,处理高维数据效果不佳,且未能关注不透水面具体材质信息。为更有效利用无人机高光谱数据获取城镇不透水面信息,评估城镇建设发展情况,选择河北省张家口市怀来县东花园镇为研究区域,从机载高光谱遥感数据中获取了150个有效波段。在此基础上,运用逐步判别分析法选择适用于城镇不透水面提取的高光谱特征波段,并使用波段标准差、波段间相关性和主成分分析的方法进行验证和综合分析,最终确定了14个具有代表性的波段。随后,提出了一种基于卷积神经网络的遥感不透水面提取方法。通过改进AlexNet网络架构,构建了一个包含四个卷积层、一个池化层和两个全连接层的深度学习网络模型。最后,在研究区设计了两组对比实验,分别比较高光谱原始影像与选取特征波段的不透水面信息提取精度,以及提出的网络模型与常见不透水面提取方法的信息提取精度。结果表明,所选的特征波段组合能够作为不透水面提取的最佳波段组合,显著提升了各类方法的提取精度。同时,提出的网络模型为不透水面提取的最优方法,结合最佳波段组合,最终分类的总体精度和Kappa系数分别达到了99.07%和0.9883,表现优异。该研究成果对于城镇建设的可持续发展和生态环境保护具有重要意义,可为相关领域的研究提供有力支持。Remote sensing images acquired by UAV-mounted hyperspectral sensors have the advantages of rich spectral information and high spatial resolution,which can provide more effective data for extracting impervious surfaces in towns and cities.However,hyperspectral images contain many bands,information redundancy increases the complexity of model training,and the volume of data space grows exponentially with data dimensions.The limited sample size will be sparsely distributed in high-dimensional space,easily leading to model overfitting.In addition,the traditional extraction method has limited feature learning capability,is ineffective in dealing with high-dimensional data,and fails to focus on the specific material information of the impervious surface.To make more effective use of UAV hyperspectral data to obtain information on impervious surfaces in towns and assess the development of town construction,this study selects Donghuayuan Town,Huailai County,Zhangjiakou City,Hebei Province,as the study area and acquires 150 effective bands from airborne hyperspectral remote sensing data.On this basis,the hyperspectral feature bands applicable to extracting impervious surfaces in towns were selected using stepwise discriminant analysis,validated,and comprehensively analyzed using principal component analysis,band standard deviation,and inter-band correlation,and 14 representative bands were finally identified.Subsequently,a remote sensing impervious surface extraction method based on a convolutional neural network was proposed.By improving the AlexNet network architecture,a deep learning network model containing four convolutional layers,one pooling layer,and two fully connected layers was constructed.Finally,two sets of comparison experiments were designed in the study area to compare the information extraction accuracy of impervious surfaces in hyperspectral raw images with selected feature bands and the information extraction accuracy of the proposed network model with common impervious surface extraction methods,respec

关 键 词：高光谱影像 无人机 不透水面提取 卷积神经网络 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]