Research Of Hyperspectral Image Classification Based On Deep Adaptive Neural Network

Hyperspectral imaging technology has advanced significantly both in theory and application in the last several years.Hyperspectral images contain not only the spectral information but also the spatial information that is helpful for achieving the state-of-the-art accuracy.Hyperspectral image classification has become a hot research field;however,the data dimension increases greatly as the spatial resolution,spectral resolution and the number of selected neighbor pixels in the HSI increase,which result in high dimension and computational complexity.The traditional hyperspectral classification algorithms like the Support Vector Machine,Logistic Regression and Artificial Neural Network can be considered as shallow classifiers.It is reported that multi-layer network can extract high-level information which is helpful for pattern recognition and classification.In this paper,the principle of hyperspectral imaging and the characteristics of hyperspectral image are analyzed,then the deep belief network is introduced for the following research.This paper presents a novel deep learning algorithm named as Deep Adaptive Neural Network for hyperspectral image classification.This method can adaptively extract high-level feature from the original data by using sparsity weights.The contribution of this paper is summarized as follows:Firstly,this paper analyzes the principle of hyperspectral imaging and the characteristics of hyperspectral images.And It is necessary to reduce the dimension of the raw hyperspectral data considering the high dimensional spatial-spectral feature.Then the deep belief network is introduced for classifying hyperspectral images and the appropriate size is explored for the Deep Belief Network model.Compared with the traditional feature extraction methods,the multi-layer network can extract high-level information from raw hyperspectral data.Secondly,there is a strong correlation between different bands.A novel Deep Group Belief Network with regularized weight-decay process is proposed to reduce the band correlation and redundancy.The proposed algorithm adaptively reduces the weights of adjacent layers associated with the redundant bands,which can reduce the impact of redundant bands to improve classification accuracy.Thirdly,it's hard to implement the deep neural network in embedded systems with limited resource.This paper proposes a hardware-oriented algorithm named as deep adaptive network to reduce the computational complexity.The proposed method can adaptively remove the majority of connection weights and retain important connections.Based on the hardware-oriented framework,the Deep Adaptive Network can robustly represent the reserved connection weights using single-bit integers which can save up to 99% memory utility without undermining classification accuracy.