Hyperspectral Image Classification Based on Spectral and Texture Features
扫码查看
点击上方二维码区域，可以放大扫码查看
原文链接
国家科技期刊平台
NETL
NSTL
万方数据
针对高光谱图像分类技术利用空间信息不足的问题,提出了一种光谱特征和纹理特征相结合的高光谱图像分类方法.通过二维 Gabor小波提取高光谱图像纹理特征,利用函数型数据分析(FDA)框架分析高光谱数据,采用三次 B 样条基系统生成光谱特征和纹理特征的函数型数据,结合函数主成分分析(FPCA)提取每个像素的函数主成分(FPC),运用概率 SVM分别对光谱特征和纹理特征进行分类.通过实验调参找到光谱特征和纹理特征的最佳参数组合,从而提高分类精度.通过在两个具有不同空间分辨率的高光谱图像数据集上进行实验,分析了参数的变化对分类精度的影响,并与其他同类方法相比较,随机选择5%的样本和 10%的样本作为训练样本的总体精确度(OA)较 EMAP+SVM 方法分别提高了 1.39%和3.87%.
A hyperspectral image classification method based on the combination of spectral features and texture features is proposed to solve the problem of using insufficient spatial information in hyperspectral image classification technology.The texture features of hyperspectral images are extracted by two-dimensional Gabor wavelet,hyperspectral data are ana-lyzed by functional data analysis(FDA)framework,functional data of spectral features and texture features are generated by cubic B-spline basis system,functional principal component analysis(FPCA)is used to extract the functional principal com-ponent(FPC)of each pixel,and probabilistic SVM is used to classify spectral features and texture features respectively.The best parameter combination of spectral features and texture features is found by adjusting parameters in experiments,so as to improve the classification accuracy.Through experiments on two hyperspectral image datasets with different spatial resolutions,the influence of parameter changes on classification accuracy is analyzed.Compared with other similar methods,the overall accuracy(OA)of randomly selecting 5%samples and 10%samples as training samples is 1.39%and 3.87%higher than that of EMAP+SVM respectively.
hyperspectral imageimage classificationtexture featuresspectral featuresspatial information
国家科技期刊平台
NSTL
万方数据
