基于多尺度3D-2D卷积神经网络的高光谱图像分类

Hyperspectral Image Classification Based on Multiscale 3D-2D Convolutional Neural Network

设计提出了1种针对高光谱图像分类任务的3D-MSCNN模型。在PCA降维的基础上,利用3D空谱特征提取网络和2D多尺度特征提取网络实现高光谱图像特征提取,充分发挥高光谱图像空谱信息价值,增强对不同尺度地表覆盖的表达能力。最后,利用Softmax分类损失函数实现高光谱图像分类任务。实验结果表明,本文算法在In⁃dian Pines和Pavia University数据集上都取得了较好的分类效果。与CD-CNN、3D-CNN、SS-Net和HybirdSN等方法相比,本文算法能够有效提升总体精度、平均精度和Kappa系数等客观评价指标。

A 3D-MSCNN model for hyperspectral image classification task is designed and proposed.On the basis of PCA dimensionality reduction,hyperspectral image feature extraction is realized by using 3D spatial-spectrum feature extraction network and 2D multi-scale feature extraction network,so as to give full play to the value of hyperspectral image spatial spectrum information and enhance the expression ability of surface coverage at different scales.Finally,the hyperspectral image classification task is realized by using softmax classification loss function.The experimental results show that the proposed algorithm has achieved good classification results on Indian pines and Pavia university datasets.Compared with CD CNN,3D-CNN,SS-NET,HybridSN and other methods,this algorithm can effectively improve the objective evaluation indexes such as overall accuracy,average accuracy and Kappa coefficients.