结合自动编码器的高光谱影像极限学习机分类

Classification of Hyperspectral Image ELM Combined with Autoencoder

极限学习机（ELM）分类作为新型神经网络算法可以实现高光谱影像的快速分类，但ELM浅层网络结构不能充分利用高光谱影像所蕴含的丰富的光谱特征。针对该问题，本文结合深度神经网络可以学习光谱深层隐含特征的优势，提出基于降噪自动编码器（DAE）的高光谱影像ELM分类方法。首先，采用DAE构造深层网络模型，利用加噪的样本数据训练网络模型，学习影像光谱的深层隐含特征；然后，用学习到的特征作为极限学习机中隐含层的输出，取代原始ELM分类中利用随机输入权值学习的浅层特征；最后，进而实现高光谱影像的分类。本文分别利用ROSIS和OMIS的高光谱影像进行分类对比实验，结果验证了该方法相对于ELM算法的分类优越性，其充分利用高光谱影像的深层光谱特征，有效提高了分类精度。

As a novel neural network algorithm, extreme learning machine （ELM） can realize rapid classification of hyper- spectral image, but the shallow network structure of ELM fails to make full use of the spectral characteristics contained in hyper- spectral image. To solve this problem, the advantage that deep neural network can learn the implicit spectral characteristics is considered, and a new ELM algorithm based on denoising autoencoders （DAE） is proposed for hyperspectral image classification. Firstly, a deep network model is constructed by DAE and trained by adding noise to sample data, and the deely implied charac- teristics of image are learned. Secondly, the learned characteristics are taken as the output of implied layer in ELM and used to replace the shallow-layer characteristics learned with random input weight in original ELM classification. Finally, the classifica- tion of hyperspectral image is realized. ROSIS and OMIS hyperspectral images are used respectively to conduct comparative exper- iment, and the result shows that the algorithm proposed in this paper is superior to ELM in classification because it can make full use of the deep spectral characteristics of hyperspectral image and effectively improve the classification accuracy.