Construction Of Generative Adversarial Network Model With Overfitting Suppression And Application Of Remote Sensing Classification

The classification of hyperspectral remote sensing image is the basis of remote sensing application,and also one of the hotspots in recent years.Hyperspectral remote sensing images usually contain hundreds of spectral bands,which provides a powerful basis for distinguishing land objects.Deep learning has powerful learning ability,which can extract the hidden features of the spectrum and classify them.It has been widely used in hyperspectral remote sensing image classification.However,the model based on deep neural network needs a large number of training samples to train a good model.Because of the high cost of tag acquisition and the small number of tag samples,training with a small number of tag samples will lead to serious model overfitting.In order to improve the classification accuracy,it is a key problem to restrain the overfitting of the deep learning model in the case of a small number of training samples.Aiming at the potential overfitting problem of the model based on deep neural network when the training samples are small,a generative countermeasure network classification algorithm with overfitting suppression is proposed and applied to hyperspectral remote sensing image classification.Firstly,the standard datasets are constructed into spatial neighborhood blocks,and the neighborhood blocks are preprocessed to avoid the interference of heterogeneous pixels in the neighborhood blocks.Then,the neighborhood blocks are divided into labeled samples,unlabeled samples and test samples,and the labeled samples and unlabeled samples are sent to the generative adversarial network for training.When input,the neighborhood blocks are trained.The pixels in the block are independently fed into the fully connected network discriminator to extract the spectral features of each pixel.Finally,the spectral features of each pixel are fused by average pooling,and the classification results are obtained by connecting them to the output layer.In each iteration process,the label information of labeled samples is used to make the real distribution of the network fitting data,and then the mean value of the high-dimensional features of the training samples is minimized.This process will update the network parameters of the discriminator,reduce the values and variances of the parameters,so as to restrain the over-fitting.The algorithm is applied to the spectral-spatial classification model designed for hyperspectral remote sensing images.The experimental results show that the overall classification accuracy reaches 89.61% and 98.79% respectively by randomly selecting 1% labeled samples from the standard datasets of Indian Pines and Pavia University.Compared with other existing algorithms,the overall classification accuracy is improved by 5.17% and 1.38% respectively.In the case of 1% labeled samples from Indian Pines datasets and 0.1% labeled samples from Pavia University datasets,the proposed algorithm outperforms several commonly used overfitting suppression algorithms,and outperforms the best Dropout algorithm.The overall classification accuracy is improved by 5.60% and 3.20%,respectively.