Hyperspectral Image Classification Using Deep Learning Method

Hyperspectral image classification has been a research hotspot in the field of remote sensing. Hyperspectral remote sensing image contains spectral information and spatial information, which facilitates the classification task. Over the past 20 years, a lot of classification and feature extraction methods have been proposed, such as nearest neighbor classifier, support vector machine classifier, linear dimension reduction based on PCA, nonlinear dimension reduction based on manifold learning. Most of these methods are in a shallow level. How to extract a deep and more abstract features to obtain a better classification performance, is an important research field of machine learning community.Deep learning is a new way to solve this problem, and also a research hotspot in recent years in the field of machine learning. Deep learning can build a deep network, increase the number of model parameters, and can remember more training data. Moreover, deep learning should construct a suitable network to solve a specific problem. With the increasing number of layers of network, the existing training methods may not be able to obtain good network, and thus new technology for training network should be used. In addition, running deep network is time consuming, and how to improve calculating efficiency is also very important This paper firstly introduces the auto-encoder network. Secondly, based on the multi-layer perceptron, it describes multiple strategies that can improve the network training efficiency and classification performance, and employed the optimized network for hyperspectral image classification. Thirdly, it utilized two deep learning methods (optimized multi-layer perceptron and convolutional neural networks) for classification of hyperspectral data by combining both of the spectral and spatial information. Finally, based on the optimized multi-layer perceptron, it construct a deep transfer network by introducing a class centroid alignment strategy for classification of multi-temporal hyperspectral data.The main work of this paper is as follows:Firstly, it makes an introduction to the stacks auto-encoder, which can well show the concept of pre-training and fine-tuning. The auto-encoder includes unsupervised pre-training to initialize the network parameter, and supervised fine-tuning to specific tasks, where the unsupervised pre-training is able to obtain better network parameters than random network initialization. However, the pre-training stage reduces the efficiency of the network. Since advanced network training techniques play important roles for the network efficieny and classification performance, we describe several effective network training methods, including ReLU neurons, methods of preventing over fitting, and optimization approaches for calculating the network parameters.Secondly, spatial knowledge of hyperspectral images is able to reduce the "pepper and salt" noise in the classification map and is quite effective in the classification task,, and therefore we introduced the spatial information into the optimized multi-layer perceptron method to classifiy hyperspectral images. In addition convolutional neural network is a very successful model for deep learning. It achieves a breakthrough in many fields, but has not been applied to hyperspectral remote sensing image classification. This paper also discusses how to use convolutional neural network to further improve the classification performance by combining with the spatial information of hyperspectral remote sensing.Finally, for classficiation of multi-temporal hyperspectral remote sensing image, if we have enough labeled samples in one image, and we can reuse the knowledge to classify new unlabeled samples of another image, the labeling cost will be greatly reduced. How to effectively use the existing knowledge to the new environment is a problem to be solved. In this paper, based on the deep learning network, the realization of transfer learning is applied to multi-temporal hyperspectral remote sensing image classification. Since the spectral features of the existing image (known as the source domain data) and the new image (known as the target domain data) may have a big change, there are problems if we directly use the labeled samples from the source domain to the target domain image classification. Therefore, data transformation should be conducted to make the two domain data become similar. This paper obtained the transformation by using a class centroid alignment strategy, and the transformed data source domain data become suitable for classification of the target domain data.The novelties of this paper include:(1) in the third chapter, some latest network optimization strategies are added into the multilayer perceptron algorithm to improve the network efficiency and classification performance. The optimized method was then applied to the classification of hyperspectral remote sensing images. (2) in the fourth chapter, convolutional neural network was applied to the classification of hyperspectral remote sensing images. Moreover, spatial information of hyperspectral images was used and two formats of network input data were designed to further improve the classification accuracy. (3) in the fifth chapter, a deep transfer network was developed based on the optimized multi-layer perceptron; the network is able to do transfer learning by utilizing a class centroid alignment strategy and was successfully applied to the classification of multi-temporal hyperspectral data.