改进的残差3D-CNN的高光谱遥感影像分类

Hyperspectral remote sensing image classification based on improved residual 3D-CNN

针对高光谱遥感影像分类中空间特征和光谱特征利用率低问题,该文综合三维卷积神经网络、谷歌神经网络和残差神经网络的优势,提出融合改进Inception模块的残差三维卷积神经网络高光谱遥感影像分类方法。改进后的Inception模块包括4条不同的卷积层分支,用以提取蕴涵在高光谱遥感影像中多尺度的特征;利用了3D卷积核代替2D卷积核能直接同时提取高光谱遥感影像中更丰富的空-谱特征;通过残差结构连接分支提取特征缓解了梯度消失的问题,提取更深层次的特征。实验表明,该文算法不仅提高了条状和线状地物区域的边缘分类准确率,对小目标的分类能力也得到了增强。

Aiming at the low utilization rate of spatial features and spectral features in hyperspectral remote sensing image classification,this paper proposed a residual three dimensional convolutional neural network,3D-CNN hyperspectral remote sensing image classification method which integrated improved Inception module,combing the advantages of three-dimensional convolutional neural network,Google neural network and residual neural network.The improved Inception module included four branches of different convolution layers to extract multi-scale features contained in hyperspectral remote sensing images.3Dconvolution kernel was used instead of 2D convolution nuclear energy to extract more space-spectral features directly and simultaneously in hyperspectral remote sensing images.By connecting branches with residual structure,the problem of gradient disappearing was alleviated and deeper features were extracted.Experiments showed that the proposed algorithm not only improved the edge classification accuracy of strip and linear feature areas,but also enhanced the classification ability of small targets.