Predicting Unknown Classes on Hyperspectral Image Data Using Deep Learning Techniques

Most of the current HSI Classification methods assumed that the dataset belongs to a closed and complete world. The possibility of the unseen data having novel or unknown classes was not considered. Two novel methods: Hybrid-MDL4OW and WPC-HSI, are proposed in this research study to predict unknown classes for HSI Data. Focus is on the extraction of a more distinctive feature space. Evaluation has been done for Imperfect Classification Systems (ICS). Almost Perfect Classification Systems (APCS) have majority of the classes present in the known sample while ICS have only few classes in the known sample. Complete ICS have all the basic land cover elements in the training sample while Incomplete ICS have either natural or impervious classes. The proposed methods significantly improve the F1 Scores for both kinds of ICS. For example, Hybrid-MDL4OW improved the F1 Score for Pavia OW2 ICS by 17.8% while WPC-HSI improved the F1 Score by 41.55% for Pavia OW1 ICS. The proposed methods are also found suitable for limited training data.