联合矢量数据和深度学习的遥感影像对象级分类样本自动选择方法  

Automatic Selection of Remote Sensing Image Object Level Classification Samples Based on Vector Data and Depth Learning

作　　者：何燕兰 王胜利 朱寿红[3] 刘文杰 HE Yanlan;WANG Shengli;ZHU Shouhong;LIU Wenjie(Jiangsu Geologic Surveying and Mapping Institute,Nangjing 211102,China;School of Environment and Spatial Informatics,China University of Mining and Technology,Xuzhou,Jiangsu 221116,China;Rand Project Land Technology Co.Ltd.in Jiangsu Province,Nanjing 210019,China)

机构地区：[1]江苏省地质测绘院,南京211102 [2]中国矿业大学环境与测绘学院,江苏徐州221116 [3]江苏省兰德土地工程技术有限公司,南京210019

出　　处：《遥感信息》2023年第6期15-21,共7页Remote Sensing Information

基　　金：江苏省地质矿产勘查局科研项目(2020KY11、2022KY15)。

摘　　要：针对目前的样本获取手段过于依赖人工制作,难以满足当前业务化实际需求的问题,提出了一种基于历史矢量数据和双线性差异化集成卷积神经网络支持的对象级样本自动选择方法。该方法首先通过对影像多尺度分割获取同质性较高的地物块状图斑,将历史矢量携带的标签信息赋值给该块状图斑;然后,通过图斑边界约束自适应生成多尺度样本集;最后,利用双线性差异化集成卷积神经网络进行样本的选择和纯化,通过属性关联实现对象级的高质量样本获取。无人机影像的分类结果表明,该方法充分结合了历史矢量数据先验几何约束和属性信息,顾及了最新影像中地物的光谱特性、边界特征和纹理信息,并引入深度学习方法实现了多尺度样本的纯化处理,实现了快速获取满足实际需求的高可靠性对象级分类样本。Aiming at the problem that the current sample acquisition method still relies on manual production and is difficult to meet the actual needs of the current business,this study proposes a method based on historical vector data and bilinear differential integrated convolutional neural network(IB-CNN)support.Firstly,the multi-scale segmentation is used to obtain patches with high homogeneity and the label information carried by the history vector is assigned to the patches.Then the multi-scale sample set is generated adaptively by patch boundary constraint.Finally,IB-CNN is used for sample selection and purification,and object-level high-quality sample acquisition is achieved through attribute association.The classification results of UAV images show that this method can make full use of the prior geometric constraints and attribute information of the historical vector,and take into the spectrum,boundary and texture information of the latest image,and introduce deep learning algorithms to purify multi-scale samples,so as to quickly obtain highly reliable object-level classification samples that meet the requirements.

关 键 词：矢量数据 双线性差异化集成卷积神经网络 多尺度样本集 面向对象 样本自动选择 

分 类 号：TP75[自动化与计算机技术—检测技术与自动化装置]