Intraclass Similarity Structure Representation-Based Hyperspectral Imagery Classification With Few Samples

Abstract

Hyperspectral imagery (HSI) classification is one of the fundamental applications in remote sensing domain, which aims at predicting the labels of unlabeled pixels in an image with a classifier trained on a certain amount of labeled pixels. However, due to the expensive cost on manual labeling, only limited labeled pixels can be obtained in real applications, which is prone to result in the training of classifier to be overfitting. To address this problem, we present an intraclass similarity structure representation-based HSI classification method. First, according to the intraclass spectrum similarity of pixels, we establish a mixed labels-based annotation model. Given some randomly selected unlabeled pixels, we employ the proposed annotation model to assign each pixel a mixed label from the top-two possible classes, and then augment the original training set with those labeled pixels. On the augmented training set, we train a deep convolutional neural network-based classification model. With several individual rounds of the annotation and classifier training procedures, we obtain several independent classification models and predict the final labels as their fusion results with a voting strategy. Experimental results demonstrate the effectiveness of the proposed method in terms of HSI classification with few training samples.