基于ML-3DCNN的高光谱遥感图像分类算法

Classification Algorithm for Hyperspectral Remote Sensing Images Based on ML-3DCNN

高光谱遥感图像含有丰富的光谱和空间特征信息,在进行地物分类时这些特征的提取尤为重要。提出了一种基于融合拉普拉斯特征的三维卷积神经网络(Meld Laplace 3D Convolution Neural Network,ML-3DCNN)的高光谱遥感图像分类算法,该算法采用主成分分析(Principal Components Analysis,PCA)降维结合空间边缘细节特征提取组成的双分支网络结构对遥感图像进行特征提取,有利于提升分类性能。在公开的3组高光谱遥感图像数据集上,与PCA+3DCNN和PCA+3D-2DCNN算法进行对比,结果表明提出的双分支网络结构提升了高光谱遥感影像的分类精度。

Hyperspectral remote sensing images contain the rich spectral and spatial feature information.Extracting appropriate features is particularly important to classify the ground objects.In this paper,a hyperspectral remote sensing image classification algorithm based on Meld Laplace 3D Convolution Neural Network(ML-3DCNN)is proposed.The algorithm adopts a two-branch network structure composed of Principal Components Analysis(PCA)dimension reduction and the spatial edge detail feature extraction,which is conducive to improving the classification performance.Compared with PCA+3DCNN and PCA+3D-2DCNN algorithms,on 3 public hyperspectral remote sensing image benchmarks,the experimental results show that the proposed double-branch network structure improves the classification accuracy of hyperspectral remote sensing image.