联合局部二值模式与K-最近邻算法的高光谱图像分类方法
投稿时间:2020-06-29  修订日期:2020-07-22  点此下载全文
引用本文:
摘要点击次数: 205
全文下载次数: 0
作者单位E-mail
赵晋陵 安徽大学 农业生态大数据分析与应用技术国家地方联合工程研究中心 aling0123@163.com 
胡磊 安徽大学 电子信息工程学院  
严豪 安徽大学 电子信息工程学院  
储国民 安徽大学 电子信息工程学院  
方艳 安徽大学 电子信息工程学院  
黄林生 安徽大学 农业生态大数据分析与应用技术国家地方联合工程研究中心  
基金项目:国家自然科学基金“基于无人机遥感影像融合的地块尺度小麦白粉病解析方法研究”(31971789);安徽省自然科学基金“小麦白粉病侵染风险评估体系构建与预测模型研究”(2008085MF184);安徽省高等学校自然科学研究项目(KJ2019A0030)
中文摘要:如何利用较少训练样本达到高分类精度已成为高光谱遥感领域的重要研究方向和极具挑战性任务。针对高光谱图像包含的丰富光谱与空间信息,提出一种联合局部二值模式LBP(Local Binary Patterns)与K-最近邻KNN(K-Nearest Neighbors)算法的高光谱图像分类方法。首先,通过主成分分析PCA(Principle Component Analysis)对高光谱数据进行降维;然后,使用LBP提取降维后的高光谱图像空间纹理信息,将光谱与空间特征变量堆叠成空—谱特征向量;最后,输入最近邻分类器得到分类结果。选取Pavia University、Indian Pines和Salinas三种公开高光谱数据集的训练集和测试集作为建模和验证数据源,选取KNN、径向基核函数支持向量机RBF-SVM(Radial Basis Function Support Vector Machine)、核联合正交匹配追踪(Kernel Simultaneous Orthogonal Matching Pursuit,KSOMP)三种经典分类算法作为比较。在Pavia University与Indian Pines数据集中随机选取10%作为训练样本,总体精度和Kappa系数分别达到99.15%、98.87%和97.88%、97.58%;在Salinas数据集中随机选取2%作为训练样本,总体精度与Kappa系数为98.46%和98.29%。实验结果表明,在训练样本仅为数据集10%甚至2%的条件下,本文提出的方法仍可达到98%以上的分类精度,可满足训练样本难以获取的应用场景对高分类精度要求。
中文关键词:高光谱遥感  局部二值模式  k-最近邻  空谱特征  主成分分析
 
Hyperspectral image classification combing local binary patterns and k-nearest neighbors algorithm
Abstract:It is a highly important and challenging task to finish the high-accuracy hyperspectral image classification using fewer training samples. A novel hyperspectral image-based classification method (hereafter referred to as the LBP-SSKNN) was proposed by combing Local Binary Patterns (LBP) and K-Nearest Neighbors (KNN). First, the Principal Component Analysis (PCA) was used to reduce the dimension of hyperspectral image. Subsequently, the LBP was used to extract the spatial texture information and the spatial and spectral features were uniformly scaled to form the spatial-spectral vectors. Finally, the vectors were input into the KNN classifier to obtain the classification result. The training and test datasets of three popular open hyperspectral datasets were used to validate the proposed method, including Pavia University, Indian Pines and Salinas. The classification method was verified on three groups of hyperspectral remote sensing image datasets. In addition, three classic classifiers were selected to compare the LBP-SSKNN, namely, KNN, Radial Basis Function Support Vector Machine (RBF-SVM) and Kernel Simultaneous Orthogonal Matching Pursuit (KSOMP). In the Pavia University dataset and Indian Pines dataset, the 10% of training dataset were randomly selected as the training samples. The overall accuracy (OA) and Kappa coefficient reach 99.15%, 98.87% and 97.88%, 97.58%, respectively. In the Salinas dataset, only the 2% of training dataset were randomly selected as the training samples, and the OA and Kappa coefficient reach 98.46% and 98.29%. The experimental results show that the OA of LBP-SSKNN method can still reach more than 98% under the 10% and even 2% of the training dataset. Our proposed method can satisfy the high-accuracy requirement due to limited training samples in practical application scenes.
keywords:hyperspectral  remote sensing, local  binary patterns (LBP), k-nearest  neighbors (KNN), spatial  and textural  features, principle  component analysis
  HTML  查看/发表评论  下载PDF阅读器

版权所有:《红外与毫米波学报》编辑部