Research Of Hyperspectral Image Classification With Small Sample Set Based On Deep Learning

With the rapid development of space-imaging technology,the hyperspectral remote images are more accessible,and widely applied in more fields.Hyperspectral images have rich spectral information and high-resolution spatial information,which makes it have great value to be used for classification of earth-features.Hyperspectral images contain not only high-dimensional data and well-defined spatial contours,but also a large amount of noise information,so extracting valuable and clear features is essential for image classification.Traditional manual feature extraction methods are time-consuming and laborious,and can only extract shallow features.As an important part of machine learning,deep learning makes great progess in image processing with its unique advantages of hidden-learning and deep feature extraction.The application of deep learning model for classification of hyperspectral remote sensing images has important value of research and reality.Firstly,in view of a lack of training samples with labels,an improved data augmentation method based on the diversity of image rotation is proposed in this paper.After the improved method,more corresponding train-samples are obtained,and the comparative experiments verify the effectiveness of the improved method.Secondly,excessive number of layers will result in the degradation of generalization ability of deep learning model.To solve this problem,a new deep learning network combined with the principle of residual learning is proposed.It is verified on three experimental data sets that the new network can achieve better performance than some state-of-the-art methods.Thirdly,the unsupervised learning of deep learning model was studied and a multi-scale filter convolutional neural network(MSF_CNN)is proposed to extract features of different spatial sizes.The MNF transform was used to construct the sample label image which is used for the learning of the MSF_CNN model.The comparative experiments verifies the effectiveness and advancement of the MSF_CNN model in feature extraction of hyperspectral images.