Research On Feature Learning And Classification Of Hyperspectral Remote Sensing Images Based On Deep Learning

Compared with the traditional remote sensing technology,hyperspectral remote sensing has many characteristics such as rich band,high resolution,large amount of data and increased information redundancy.It not only provides 2D spacial images,but also the spectral information.In recent years,hyperspectral remote sensing has been highly valued by domestic and foreign scientists.Hyperspectral remote sensing image classification as one of the important ways to obtain information has become a hot topic in the field of remote sensing.There are few labeled samples in hyperspectral remote sensing images but the cost of manual labeling is high.These lead to over-fitting and low classification accuracy in traditional methods based on neural network.In order to deal with this problem,we propose a hyperspectral image classification method based on Auxiliary Classifier Generative Adversarial Networks(HS-ACGAN).The advantage of Generative Adversarial Network is used to reduce the demand for labeled samples,at the same time,the spectral features are learned to classify the samples.We also investigated the effect of Training-to-Total sample Ratio(TTR)and Batchsize two parameters on classification accuracy.The hyperspectral image is susceptible to noise and have the phenomena of“same object with different spectrums” and “different objects same image”.In order to solve these problems,we present a spectral feature classification model based on Auxiliary Classifier Generative Adversarial Networks(ACGAN-CNN).We treat the pre-trained discriminator in Auxiliary Classifier Generative Adversarial Networks as aspectral feature extractors.Then,use CNN to classify spectral features.Experiments show that this method can reduce misclassification and improve classification accuracy.There are few classification methods based on Generative Adversarial Networks,but only a few use the spatial information of the data.This paper proposes a spectral-spatial feature extraction and classification method based on ACGAN(ACGAN-LBP-CNN).First,the spectral features of the sample were extracted by the pre-trained ACGAN model.Second,we calculate the entropy of each band for band selection,which helps to avoid redundancy and reduce computation.Thirdly,we use the rotation invariant LBP to extract the texture features.Finally,we fuse spectral features and spatial features,and then use convolution neural networks to train and classify the fusion features.Experiments and comparisons on two datasets,Indian Pines and Pavia University,show that the method proposed in this paper requires fewer training samples but get higher classification accuracy.The integration of spatial features improves the misclassification caused by a single feature,reduces the noise in the results,and improves the classification accuracy.