Researches On Hyperspectral Images Classification By Gabor Filtering Based Deep Network

Hyperspectral images have rich and reliable spectral information.Hence,hyperspectral images play an important role in earth observation system and are widely used in geophysics exploration,agricultural remote sensing,ocean remote sensing and environmental monitoring.In the study of hyperspectral image classification,many spectral-spatial classification methods have been proposed.Recently,deep learning has been a very active research topic in the field of image processing.Compared with the traditional feature extraction methods,multi-layer deep learning network can extract the more abstract high-level features through the combination of low-level features.The abstract high-level features show the nonlinear distributed feature of the data,which have superior discriminability and robustness in image processing.In this paper,effective spatial structure features of hyperspectral images are extracted,and the deep features are captured by training the deep network with the fused spectral-spatial features for classification.Thus,the method of hyperspectral images classification by Gabor filtering based deep network is proposed.Experiments on three widely used real hyperspectral data sets show that the proposed method outperforms several well-known classification methods in terms of classification accuracy.The main contents of this thesis are as follows:1.The method of hyperspectral images classification by Gabor filtering based sparse auto-encoder deep network.Spatial features of hyperspectral image are extracted via Gabor filtering.Gabor filter can capture physical structures of hyperspectral images,such as texture and direction information.High-level features are learnt by a stacked sparse auto-encoder deep network with fused spectral and Gabor features.Since the number of reference training samples of hyperspectral images is often very limited,which negatively affects the classification performance in deep learning,an effective way of constructing virtual samples is designed to generate more training samples,automatically.By jointly utilizing both the real and virtual samples,the parameters of the deep network can be better trained and updated,which can result in more robust and accurate classification results.2.The method of hyperspectral images classification by Gabor filtering based deep belief network.Gabor features and spectral features are still staked to form combined features.The parameters of deep belief network are captured by trained deep belief network with the fused features.The parameters of deep belief network are used as the initialization parameter of BP neural network.The combined features make full use of structure information of hyperspectral,and result in that parameters of deep network can be better learnt and updated.Meanwhile,BP neural network that is initialized by the parameters of deep belief network can be provided a prior knowledge in the training process,which increases the discriminability of the network.3.A software system is designed for hyperspectral image classification mentioned above.The software system is programmed based on MATLAB 2014a programming platform by GUI application framework.The software system can achieve the method of hyperspectral images classification by Gabor filtering based sparse auto-encoder deep network and deep belief network.