利用深度特征融合进行高光谱遥感影像分类

Hyperspectral Remote Sensing Image Classification Using Depth Feature Fusion

针对卷积神经网络(convolution neural network,CNN)未能充分挖掘和利用遥感影像不同层次空-谱特征信息的问题,提出了一种多输入中高层特征信息融合方法对高光谱遥感影像进行分类。文章利用三维卷积神经网络(three dimensional convolution neural network,3D-CNN)的“立体”感受野,深入挖掘高光谱影像的空-谱联合特征信息,分析深度网络各层特征信息对图像分类的影响,提出优化的特征融合策略,并利用“珠海一号”卫星拍摄的高光谱影像对3个地区进行地物分类实验。实验结果表明,由于充分利用了高光谱影像的空-谱联合特征,通过特征融合集成了不同网络层的特征优势,相对于其他高光谱分类方法,该方法具有更好的分类效果。

Aiming at the problem that convolution neural network(CNN)cannot fully extract and utilize the spatial spectrum feature information of each layer of remote sensing imagery,a multi-input feature fusion method for hyperspectral remote sensing image classification is proposed.In this paper,the spatial and spectral joint feature information of hyperspectral images is deeply excavated by using the“three-dimensional”receiver field of three dimensional convolution neural network(3D-CNN),the influence of feature information of each layer of the deep network on image classification is analyzed,and an optimized feature fusion strategy is proposed.The hyperspectral image taken by“Zhuhai No.1”satellite is used to carry out the classification experiment of three areas.The experimental results show that the proposed method has a better classification effect than other hyperspectral classification methods because it makes full use of the space-spectrum joint features of hyperspectral images and integrates the feature advantages of different network layers through feature fusion.