Unsupervised Hyperspectral Image Clustering with Interval Type-2 Uncertainty Fuzzy Modelling Approaches

In recent years, hyperspectral imaging (HSI) has proved to be immensely helpful in material identification and environmental monitoring, due to its ability to capture spectral signatures across different wavelength bands. However, due to high dimensionality and variability of spectral signatures, it is challenging to effectively implement unsupervised learning models – especially clustering methods which are primarily used for defining HSI class boundaries. The study begins by examining various methods that are currently used to process hyperspectral images and label their pixels. Although many techniques such as support vector machines, neural networks, K-means – mitigate issues like noise, redundant bands, limited labels, and some are even able to handle high dimensionality, mixed‐pixel ambiguity and spectral variability at class boundaries still pose significant challenges and leave much to be desired. With this thesis research, we attempted to address these limitations, utilizing fuzzy uncertainty-based approaches such as traditional fuzzy sets, type-2 fuzzy sets (T2 FSs) and related Fuzzy C-Means (FCM) clustering. With T2 FSs, type-reduction is performed using two currently known methods: Nie-Tan and Karnik-Mendel, yielding crisp values of the final-cluster centers. Advancing from available unsupervised clustering approach, we introduce an interval type-2 fuzzy c-multiple means (IT2-FCMM). The proposed approach is particularly relevant for applications where class boundaries are in the “grey area”—such as agricultural crop mapping (e.g., distinguishing mixed-crop types, detecting signs of disease in specific patches), mineral exploration (e.g., identifying mixed assemblages where two or more minerals grow together) and environmental change detection (e.g., monitoring deforestation over time, tracking shrinking of wetlands)—since it provides both accurate cluster assignments and a measure of confidence for each pixel. The algorithm has been implemented in Python and applied to publicly available benchmark hyperspectral datasets. Comparative experiments with other methods demonstrate that the proposed method achieves greater robustness to spectral variability and noise than currently known FCM and other centroid-based clustering approaches. Computation has been accelerated via GPU implementation.