Semi-Supervised Hyperspectral Image Classification Based On Deep Learning

Hyperspectral image classification is an important branch of remote sensing hyperspectral image processing.Its purpose is to determine the true labels of each pixel in the remote sensing hyperspectral image by computer.With the great success of deep learning in the fields of machine learning and pattern recognition,many scholars have applied deep learning to the classification of hyperspectral images and achieved good results.Labeling the training samples required for hyperspectral image classification is time-consuming and labor-intensive,because they have a large amount of data and it's difficult to label each pixel of the data.Therefore,using a large number of unlabeled samples and a small number of labeled samples to classify hyperspectral images is a very meaningful thing.Based on the above analysis,this paper designs several hyperspectral image classification frameworks based on semi-supervised deep learning to classify hyperspectral images,so as to solve the problem that hyperspectral image is difficult to label and deep learning requires a large amount of labeled data.The main contents are as follows:Firstly,a semi-supervised neural network model capable of training with a small number of labeled tags and a large number of unlabeled hyperspectral pixels was designed.In traditional deep learning,unlabeled samples were typically used to pre-train the network layer by layer to initialize the network.In the network,it can perform reconstruction tasks by adding a decoding channel to the network while performing classification tasks,and the features of the top layer can be focused on classification tasks by adding a horizontal connection between the encoder and the decoder,The features of original encoder can be reconstructed directly through these lateral connections.Therefore,the unlabeled sample can assist the labeled sample,and finally achieved a good classification effect in the experiment.Secondly,a network that can combine the spatial-spectral information of hyperspectral images with semi-supervised classification is designed.The spatial information of hyperspectral images is ignored in the last method,and the existing methods combining spatial information and deep learning can only use labeled samples for training.Therefore,this paper designs a semi-supervised 3D convolutional neural network capable of extracting spectral-spatial information of hyperspectral data simultaneously.The key of this model is to extract features by combining combine the spatial-spectral information,and at the same time make full use of a large number of unlabeled samples,through which a better classification effect can be obtained.Finally,a relational network capable of supervised learning the relationship between two hyperspectral pixels is designed,and a method for pseudo-labeling of unlabeled samples is designed by using the metric relationship learned by the relational network.The traditional unsupervised hyperspectral image metrics can't measure the relationship between two pixels very well.Therefore,this paper uses the convolutional neural network to extract the features of two pixels,and cascades the two features and then uses a new network learns the relationship between them,and finally gets a network that can distinguish whether two pixels belong to the same class.Combine this network with neighborhood information,pseudo-label a large number of unlabeled samples,and finally train a new classifier using these tagged data.