Research On Hyperspectral Image Classification Based On Multiscale And Multiscene Transfer Learning

Hyperspectral remote sensing has higher spectral resolution,which makes people's cognition of features in objects more intensive.As an important part of hyperspectral remote sensing,classification methods determine the production and the application of the following thematic maps to a large extent.However,the contradiction between the rich spectral-spatial information of hyperspectral images and the lack of labeled samples severely restricts the development of hyperspectral image classification techniques.Therefore,based on the characteristics of hyperspectral imagery,this thesis analyzes and studies the effective extraction of spectral-spatial information and the adaptability of samples in multiple scenes.Firstly,in order to realize the effective extraction of the spectral-spatial information,an adaptive bilateral filter by improving the joint bilateral edgepreserving filter is proposed to extract the multiscale spectral-spatial features,which avoids the complex parameter tuning of the joint bilateral edge-preserving filter and improves the classification performance.In addition,for the problem of “salt and pepper noise” which is easy to occur on the pixel-level classification maps,a post-processing method based on the probability maps is introduced.By strengthening the correlation between pixel points,the continuity on the spatial region of the classification maps is improved and the further improvement in classification accuracy is achieved.Secondly,the thesis further studies the problem of spectral-spatial feature extraction and sample deficiency in complex scenes.First,based on the deep learning theory,a Multiscale Spectral-Spatial Unified Network(MSSN)is proposed.With the better feature extraction capability of the deep network,the classification performance in complex scenarios is effectively enhanced.Then,because the phenomenon of lack of labeled samples is more prominent in deep networks,the semi-supervised learning method based on “pseudo-labels” is introduced for it.The method makes full use of a large number of unlabeled samples,and improves the classification performance of MSSN network when the training samples are insufficient.The experimental results also prove the effectiveness of the method.Finally,based on MSSN network,the thesis studies transfer learning on the problem in terms of the adaptability of samples in multiple scenarios.The modelbased transfer learning method and the feature-model-based transfer learning method are respectively to solve the network transfer in multiple images achieved by same and different hyperspectral sensors,which the number of samples in the new scene image is increased the applicability of the samples is enhanced.What's more,the methods effectively solve the problem of insufficiency in labeled samples and can help to improve the classification performance.