多层级二值模式下的卷积网络高光谱影像分类
Hyperspectral Image Classification Based on Convolutional Neural Network and Multi-layer Binary Pattern
职露 1胡涛 2尹宾宾 3余旭初 2王彦坤4
作者信息
- 1. 郑州师范学院地理与旅游学院,河南郑州,450000;信息工程大学数据与目标工程学院,河南郑州,450000
- 2. 信息工程大学数据与目标工程学院,河南郑州,450000
- 3. 天津飞眼无人机科技有限公司,天津,300000
- 4. 深圳职业技术大学物联网研究院,广东深圳,518000
- 折叠
摘要
针对如何改善小样本下卷积神经网络(convolution-al neural network,CNN)的高光谱影像分类效果,结合人工设计特征,提出多层级二值模式下的卷积网络高光谱影像分类方法.该方法先采用多层级二值模式进行高光谱影像纹理特征提取,从不同尺度反映影像的局部细节信息,生成更具鉴别性的特征.以此纹理先验特征为基础,利用卷积神经网络对其进行进一步的自动学习与分类.为验证该分类方法的有效性,选取空间分辨率、地物覆盖类型不同的PaviaU和Salinas高光谱影像实验数据,分别对局部二值模式、多层级二值模式、Gabor、GLCM(gray-level co-occurrence matrix)进行特征判别能力分析,并针对各特征开展卷积网络分类实验.结果表明,多层级二值模式下卷积网络分类总体分类精度分别达到91.98％、98.15％,优于纯光谱、Gabor等分类.
Abstract
The hyperspectral image classification method based on convolutional neural network and multi-layer binary pattern is proposed to improve the classification effect of hy-perspectral image of convolutional neural network with small samples via the artificial designed features. Firstly,the textur-al features are expressed using multi-layer binary pattern re-flecting the local details from different scales. On this basis,the deeper automatic learning and classification are carried out using convolution neural network. In order to verify the effec-tiveness of the proposed method,PaviaU and Salinas with dif-ferent spatial resolution and ground cover are used. Five kinds of features such as the spectra,the local binary pattern,Ga-bor,etc.,are employed for feature discriminative ability anal-ysis and hyperspectral image classification. The overall classi-fication accuracies with the proposed method respectively reach 91.98％ and 98.15％,which is superior than the other methods.
关键词
高光谱影像分类/卷积神经网络/纹理特征/多层级二值模式/局部二值模式Key words
hyperspectral image classification/convolutional neural network/textural feature/multi-layer binary pattern/local binary pattern引用本文复制引用
基金项目
国家自然科学基金(31971723)
国家自然科学基金青年基金(42001389)
福建省自然资源厅科技创新项目(KY-080000-04-2021-030)
出版年
2024
NETL
NSTL
万方数据