一种高光谱遥感影像的三维卷积神经网络语义分割方法

A semantic segmentation method of hyperspectral remote sensing image based on three-dimensional convolution neural network

高光谱图像具有丰富的光谱信息和空间信息,综合运用空间特征和光谱特征是提高高光谱图像分类精度的关键.针对传统二维的卷积神经网络无法充分利用高光谱丰富光谱信息的问题,设计一种基于三维卷积神经网络(Three-dimensional convolutional neural network, 3D-CNN)的深度卷积编解码网络,将三维卷积神经网络引入编码结构,同时提取光谱和空间特征,并且在池化层引入池化索引策略;解码部分利用最大池化索引上采样操作.两个高光谱遥感影像公开数据集的分类实验结果表明,实现了高光谱的空间和光谱特征的融合提取,较基于2D-CNN的分类方法能够获得更高的分类精度.

Hyperspectral images are rich both in spectral and spatial information.The comprehensive use of spatial and spectral features is the key to improve the classification accuracy of hyperspectral images.An improved three-dimensional convolutional neural network(3D-CNN) with self-coding and decoding structure is developed in order to solve the problems of classification for Hyperspectral images.In the coding part, the three-dimensional convolution neural network is introduced into the coding structure, so that the convolution is carried out simultaneously in the three dimensions of spectrum and space.The pooling index strategy of classical 2D-CNN network is introduced in the pooling layer.The decoding part adopts the corresponding transpose 3D-CNN and the maximum pool index is used to simplify the up-sampling operation.Experiments show that this method realizes the fusion and extraction of hyperspectral spatial and spectral features for classification with hyperspectral remote sensing images.The experimental results on two public data sets of hyperspectral remote sensing images show that the classification accuracy of hyperspectral images with the proposed method owns higher efficiency and higher accuracy than that of two-dimensional convolution network.