Hyperspectral Images Classification Based On Deep Random Forest And Knowledge Distillation

Hyperspectral images are typically composed of hundreds of high-resolution spectra,with richer information content and more powerful feature representation than multi-spectral images.This allows for more accurate results when performing feature classification processing on hyperspectral images.In recent years,deep learning has been applied to various fields as an emerging computational method,and has achieved good results in hyperspectral image classification.This thesis mainly uses the deep convolutional neural network to classify hyperspectral images and uses the knowledge distillation method to make a network with a small amount of parameters achieve satisfactory accuracy.Deep networks often have large network specifications while achieving high accuracy.However,excessive network scale is not allowed due to various conditions in practical applications.But if we use the small network directly,it is difficult to obtain a good classification effect due to the singularity of the information contained in the label and the limited ability of the network expression.Therefore,we adopt the basic idea of the knowledge distillation method,using the larger network to generate the prediction probability including some information,and use it as a training label to guide another smaller network,so that the smaller network could achieve substantially the same accuracy.Because the actual effect of the original method is not ideal,we proposes a different way to learn about the knowledge of the big network.We make use of a large number of virtual samples to transfer the the knowledge of the big network and the influence of the actual methods on the effects of different ways or quantities of virtual samples is explored to provide guidance on the practical application of our proposed method.In addition,the traditional deep neural network often requires a large number of training samples and a long training process.But the number of samples of hyperspectral images is often limited.For this case,the application of deep random forest model in hyperspectral image classification is studied in this paper,proposing two methods for applying the deep random forest model to the classification of hyperspectral images.One is taking use the way of dense connection between layers,and the other is fusion of multiple features.Both of the methods incorporate spatial information in order to improve the accuracy of hyperspectral image classification.Compared with the traditional neural network model,the depep random forest model has obvious advantages in training time,which can shorten the training process of hyperspectral image classification.In this paper,experiments are conducted on commonly used hyperspectral public data sets and the results fully demonstrate the effectiveness of these two methods in hyperspectral image classification.