两阶段密集特征学习的高光谱图像分类方法

Hyperspectral Image Classification Based on Two Stage Dense Feature Learning

基于卷积神经网络(CNN)的高光谱图像分类方法不断取得进展,而谱-空特征的高复杂性及低效描述仍是抑制高光谱分类精度提升的主要屏障。针对该问题,提出了一种基于密集型(Dense)谱-空特征挖掘的两阶段集成高光谱图像分类学习框架,在第一阶段进行显著波段选择,分别构建2D及3D Dense深层网络以提升高光谱数据的空间与谱间特征表达能力,在第二阶段构建分类CNN网络将融合之后的谱-空特征进一步挖掘,以提高分类特征的精细化程度。整个集成学习网络基于交叉熵损失函数进行训练学习,同时利用全连接网络构建了一种基于波段相关性的显著性波段选择方法以降低训练过程中的谱-空数据复杂度。在Purdue和KSC数据集上的实验结果表明:此方法对比其他方法的分类精度有较大提升。

Hyperspectral image classification methods based on Convolutional Neural Network continue to make progress in recent years.However,high complexity and inefficient description of spectral-spatial features can still be the main obstacle to hindering the improvement of the classification accuracy of HSIC networks.In view of this problem,a two-stage integrated learning framework based on dense spectral-spatial feature extraction is proposed.In the first stage,significant band selection is carried out initially,and then 2D and 3D dense deep networks are constructed respectively to improve the spatial and spectral featuresexpressiveness of hyperspectral data.Then,in the second stage,a classification CNN network is built to further extract the fused spectral-spatial features to improve the refinement degree of classification features,and the entire integrated learning network is trained based on the cross-entropy loss function.Meanwhile,a fully connected network is utilized to propose a salient band selection method based on band correlation,which can reduce the complexity of features.The experimental results in Purdue and KSC data sets demonstrate that this architecture has great advantages compared with other network models.