基于深度神经网络的高光谱遥感影像分类方法研究
Research on Hyperspectral Remote Sensing Image Classification Based on Deep Neural Networks
朱江 1胡华全 2范雯琦3
作者信息
- 1. 装备学院 研究生管理大队, 北京 101416
- 2. 装备学院 复杂电子系统仿真实验室, 北京 101416
- 3. 北京跟踪与通信技术研究所, 北京 100094
- 折叠
摘要
分类是高光谱遥感影像处理中最为重要的一部分.针对现有影像分类方法存在预处理复杂、高维特征提取困难、分类不够精确等缺陷,提出一种基于深度神经网络的高光谱遥感影像分类算法.算法首先采用最大噪声分数来降低特征空间维度,然后将自动编码器与softmax多项逻辑回归分类器组合成含有多隐藏层的神经网络,对高光谱遥感影像进行非监督型深度特征提取与分类.实验结果表明:与传统的基于线性支持向量机分类方法相比,本算法可提取更高级的表达特征,并在较短的处理时间内实现较好的影像分类精度.
Abstract
Classification is the most important part of hyperspectral remote sensing image processing.With view to the setbacks of complex preprocessing, difficult extraction of high-dimensional features and not accurate classification existing in the current classification method, the paper raises an algorithm of hyperspectral remote sensing image classification based on deep neural networks.This algorithm first uses the maximum noise fraction to reduce the feature space dimension, then combines the auto-encoders and the multinomial logistic regression classifier sofemax into a neural networks with multiple hidden layers, to extract and classify the unsupervised deep features of hyperspectral remote sensing images.Experiments show that, in comparison with the traditional linear support vector machine (SVM) classification method, it can extract more advanced expression features and realize better image classification accuracy in a shorter processing time.
关键词
高光谱遥感/影像分类/深层神经网络/自动编码器/最大噪声分数Key words
hyperspectral remote sensing/image classification/deep neural network/auto-encoder/maximum noise fraction引用本文复制引用
出版年
2017
NETL
NSTL
维普
万方数据