Improved hyperspectral classification of vegetation through generative deep learning models.

Abstract

Early studies into hyperspectral reflectance demonstrated that the spectra of different plants have the potential for taxonomic discrimination and classification, though this came with the caveat that misidentification was a frequent impediment as a result of small sample sizes, inter-class similarity and intra-class variability. The aim of this thesis was to develop methods of improving the ratio between intra and inter-class variability in hyperspectral vegetation spectra, and ultimately increasing classification accuracy, reliability and generalisability. This was addressed in three ways: (1) reviewing the hyperspectral classification literature of the past two decades, while also performing pre-processing and classification trials on a vegetation hyperspectral dataset, (2) increasing the number and distribution of classifier training samples through data augmentation, and (3) improving the intra/inter-class relationship with deep generative sample transformation. For objective one the last two decades of hyperspectral vegetation classification literature was systematically reviewed, specifically focusing on waveband/feature selection. Additionally, waveband selection trials were performed on a curated hyperspectral dataset in order to test the findings of the review. Both the review and waveband selection trials indicated that all characteristics of hyperspectral plant studies influence the wavebands selected for classification. However, the considerable variability in waveband selection caused by the chosen feature selection method effectively masked analysis of any variability in waveband selection caused by other aspects of the studies in the review. For this reason caution is suggested in relying upon waveband recommendations from the literature to guide waveband selections or classifications for new plant discrimination applications. As such recommendations appear to be weakly generalizable between studies. The data augmentation performed for objective two was realised through the use of a generative adversarial network (GAN), a type of generative deep learning model that could produce realistic synthetic hyperspectral vegetation data. After being trained on vegetation spectra the GAN was able to generate synthetic samples that visually matched the training spectra as well as statistically matched the distribution of each vegetation class in the training data. Creation of an augmented dataset consisting of synthetic and original samples produced training datasets with far greater sample sizes. Under almost all circumstances increases in classification accuracy of multiple classifiers was seen following their training with the augmented dataset. Objective three expanded upon the data augmentation abilities of the GAN used in objective two, introducing the ability to replicate sample spectra whilst transforming them based upon the learned features of the other classes in the study. These transformations were performed to manipulate the intra/inter-class relationships in a desired manner. This was performed on both the training and evaluation subsets of a hyperspectral vegetation dataset producing n transformed replicates of every sample where n is the number of classes in the study. Training and then evaluating the accuracy of each of these transformed subsets with multiple classification methods produced accuracies significantly higher than that of the original dataset. This significant increase in accuracy was then further improved following the ensembling of the n classification results. Visualisation of the samples used in the ensembled classification following projection to 2d space showed samples to be tightly clustered by class, indicating the successful reduction of intra-class variance as well the reduction in inter-class overlap. This thesis represents a significant step towards eliminating the lack of generalisability and transferability of vegetation classification models resulting from the pronounced effects of intra-class variance and inter-class similarity. It presents the opportunity for remote sensing practitioners to deploy their classification models to greater spatial and temporal extents whilst giving extra utility to hyperspectral samples contained within spectral libraries.