Hyperspectral Image Classification With Deep Metric Learning And Forward Propagation

Hyperspectral Image Classification is one of the relatively important research disciplines in the field of hyperspectral image processing.The main purpose of hyperspectral image classification is to accurately and efficiently classify features of interest in a hyperspectral image.With the progress of current remote sensing technology,hyperspectral image classification has been widely used in technology production and life.For example,we are familiar with land resource surveys,natural disasters,and health care.The basis of this widespread application is that humans can more easily acquire hyperspectral images with spectral resolutions up to the nanometer level.Of course,with the further improvement of remote sensing technology and spectral resolution,humans have become more and more demanding on the classification of specific feature categories under practical application scenarios.However,due to some shortcomings of the hyperspectral image itself,ssuch as the high dimensionality of the data samples caused by the rich spectral information,the large number of acquired data samples are not well calibrated and the available mark samples are limited.At present,it still seems that there are still many challenges in the classification of features in this field.With the rise of deep learning technology,more and more deep learning algorithms have also been used in this field.Although deep learning has made great progress in the field of machine vision,there are still some problems directly in the field of hyperspectral imagery.For example,there are not enough training samples to make the model easy to become over-fitting,parameter tuning is cumbersome,and so on.This paper mainly combines traditional learning methods and deep learning methods,and proposes corresponding algorithms that are more in line with hyperspectral image classification.The main work of this paper is as follows:Firstly,combining traditional metric learning thoughts with deep learning,a spectral space depth metric learning method for hyperspectral image classification is proposed.Compared with the traditional linear metric algorithm,deep metric learning can map the samples into nonlinear space by multi-layers nonlinear mapping.In the proposed model,two similar subnetworks are adopted to exploit the spectral information and spatial information,respectively.For spectral subnetwork,the original spectral features of a pixel pair are used as the input.The pixel pair can be formed by two pixels from the same class or different class.For the spatial subnetwork,the filtered spectral features of a pixel pair are used as the input.Because the local smoothness of hyperspectral imagery,the neighboring pixels often share thesame label with high probability,the low pass filtering can increase the similarity between the same class and increase the difference between the different class effectively.So,the low pass filtering is adopted to introduce the spatial information in the proposed model.Secondly,based on the integration strategy of extreme learning machine,a multi-scale deep extreme learning machine based hyperspectral image classification is proposed.The algorithm uses the traditional classifier of the extreme learning machine as the basic unit to make full use of the information between the neighborhoods.Compared with the traditional deep learning algorithm,the parameters of the algorithm are reduced,and most of the parameter verification time is omitted.By deepening the number of network layers of this algorithm,the features of each layer are fully extracted.It makes the important features in the hyperspectral data fully available,and the experimental results can achieve the desired results.Finally,the hyperspectral image classification of U-shaped networks based on random convolution is proposed.This idea is compared to the existing deep learning framework,which has a simpler architecture and takes less time.The application of random convolution kernels in hyperspectral image classification can reveal the potential utilization value of random patches and facilitate the classification of hyperspectral images.In addition,unlike the traditional deep learning method that uses only the deepest features to complete the classification,the proposed method combines shallow and deep convolutional layers in the classification,which has the advantage of multi-scale.