多层局部感知卷积神经网络的高光谱图像分类

Hyperspectral image classification of multi-layer local perceptual convolutional neural networks

摘要:

针对高光谱图像分类中光谱特征的高度非线性问题，本文提出一种基于多层感知器卷积层和批标准化层的改进卷积神经网络模型，提高模型在光谱域处理的非线性特征提取能力.该算法通过构建七层网络结构，实现多层局部感知结构，逐个像素对光谱信息开展分析，区分不同目标物的光谱信息，将全光谱段集合作为输入，舍去空间信息，利用动量梯度下降训练算法对多层局部感知卷积神经网络训练，实现对不同目标物体光谱特征的提取与分类.实验中采用两组高光谱遥感影像进行对比分析，以Pavia University数据集为例，在3600个训练样本情况下，测试集为1800个样本，本文方法正确率为90.23%，LeNet-5正确率为87.94%，Linear-SVM正确率为90.00%；在21000个训练样本情况下，测试集为全部样本，本文方法正确率为97.23%，LeNet-5正确率为96.64%，Linear-SVM正确率为92.40%.实验结果表明，在训练集较小的情况下，本文方法优于传统神经网络，能有效提取数据特征，并且在精度上和计算成本上略优于在小样本分类中具有高效和鲁棒性良好的SVM算法.在大规模训练集时，本文方法表现出良好的学习能力，在分类精度上优于LeNet-5.本文提出的多层局部感知网络结构增强了对非线性特征的学习能力，无论训练集规模大小，都比传统的SVM和一般的深度学习网络更能有效的利用高光谱图像中的逐像素点的光谱域信息，能有效提高分类精度.

Abstract:

To solve the problems on high nonlinearity in spectral features in hyperspectral image classification， a classification algorithm， based on multilayer perceptron convolutional layer and batch normalization layer improved convolutional neural network in spectral domain processing， is proposed to improve the nonlinear feature extraction ability. By constructing a sevenlayer network structure， the algorithm implements a multilayer local sensing structure， analyzes the spectral information pixel by pixel， distinguishes the spectral information of different targets， takes the full spectrum segment set as input， discards the spatial information， and uses the momentum gradient descent training. The algorithm trains multilayer local perceptual convolutional neural networks to realize the extraction and classification of spectral features of different target objects. In the experiment， two sets of hyperspectral remote sensing images are used for comparative analysis. Taking the Pavia University data set as an example， in the case of 3 600 training samples， the test set is 1 800 samples， the accuracy of the proposed method is 9023% and the accuracy of the LeNet5 and LinearSVM method are 8794% and 9000% respectively. In the case of 21 000 training samples， the test set is all samples， the accuracy is 9723%， 9664% and 9240% respectively for the proposed method， LeNet5 and LinearSVM method. The experimental results show that the proposed method is superior to the traditional neural network in the case of small training set， which can effectively extract the data features， and is superior to SVM algorithm for the small sample classification in terms of accuracy and computational cost. In the largescale training set， this method shows good learning ability and is superior to LeNet5 in classification accuracy. The multilayer local perceptual network structure proposed in this paper enhances the learning ability of nonlinear features， it can utilize hyperspectral images much more effectively than traditional SVM and general deep learning networks， both in small sample classification and large sample classification. The spectral domain information of the pixelbypixel point can effectively improve the classification accuracy.