DEEP SELF-SUPERVISED BAND-LEVEL LEARNING FOR HYPERPSPECTRAL CLASSIFICATION

Hyperspectral image classification is one of the most researched topics within hyperspectral analysis. Its importance is determined by its immediate outcome, a classified image used for planning and decision-making processes within a variety of engineering and scientific disciplines. Within the last few years, researchers have solved this task employing self-supervised learning to learn robust feature representations to alleviate the dependency on large amounts of labels required by supervised deep learning. Aiming to learn representations for hyperspectral classification purposes, several of these works use dimensionality reduction that could exclude relevant information during feature learning. Moreover, they are based on contrastive instance learning that requires a large memory bank to store the result of pairwise feature discriminations, which represents a computational hurdle. To overcome these challenges, the current approach performs self-supervised cluster assignments between sets of contiguous bands to learn semantically meaningful representations that accurately contribute to solving the hyperspectral classification task with fewer labels. The approach starts with the pre-processing of the data for self-supervised learning purposes. Subsequently, the self-supervised band-level learning phase takes the preprocessed image patches to learn relevant feature representations. Afterwards, the classification step uses the previously learned encoder model and turns it into a pixel classifier to execute the classification with fewer labels than awaited. Lastly, the validation makes use of the kappa coefficient, and the overall and average accuracy as well-established metrics for assessing classification results. The method employs two benchmark datasets for evaluation. Experimental results show that the classification quality of the proposed method surpasses supervised learning and contrastive instance learning-based methods for the majority of the studied data partition levels. The construction of the most adequate set of augmentations for hyperspectral imagery also indicated the potential of the results to further improve.