Search

In recent years, deep learning has developed rapidly thanks to advances in computing capabilities. Researchers in many fields, including remote sensing, have used this technology to achieve remarkable breakthroughs. For deep learning-based image classification tasks, including hyperspectral image classification, a large amount of labelled data is needed. However, manually labelling hyperspectral images is expensive and difficult. We aim to design a new approach to considerably improve the classification performance when the labelled data are extremely few without adding too much computational overhead. As a simple and viable solution, applying data augmentation at the pre-processing stage to improve the classifier's performance is becoming a trend. Apart from traditional data augmentation, like flipping, rotation, and adding noise, several new augmentation methods are proposed, including Cut-out, CutMix, sample pairing, and utilizing unlabelled hyperspectral data. Furthermore, existing studies have proven their effectiveness under a sparse data regime. However, when extremely few labelled samples are available, the classification accuracy of current methods is relatively low. Therefore, we propose Random Augmentation Pipe by sequentially employing a series of augmentations at the training stage. To prevent excessive computational cost, one affine transformation is introduced as a substitute for all the geometric augmentations. Moreover, the Pipe is extended by introducing five other augmentations. Among the augmentations, Square Rotation provides a new distortion without changing the original data, that effectively enhances performance. Furthermore, a principal component analysis and extended morphological profile are applied at the pre-processing stage. The Pipe is then embedded at the beginning of a carefully designed deep learning network built on convolutional neural networks and deep residual networks. Tests were performed over five commonly used datasets, including Pavia University, Pavia Centre, Botswana, Indian Pines, and Kennedy Space Centre. Through a set of comprehensive experiments, it was shown that the proposed Extended-RAP outperformed the 16 other comparison methods. The proposed methods randomly select augmentations and their magnitudes for each training iteration. Here, thirteen augmentations, including a novel rotation that was designed for hyperspectral image patches, were carefully selected and investigated. The process of multiple augmentations is regarded as the first layer of the proposed deep learning network. The layer performs augmentations during training and can be easily integrated in other models. Extended morphological profiles are applied after principal component analysis at the beginning of the pre-processing stage. To reduce the computational cost, the affine transformation is introduced as a substitute for all the geometric augmentations. [ABSTRACT FROM AUTHOR]