Clustering-Based Class Hierarchy Modeling for Semantic Segmentation Using Remotely Sensed Imagery

Land use/land cover (LULC) nomenclature is commonly organized as a tree-like hierarchy, contributing to hierarchical LULC mapping. The hierarchical structure is typically defined by considering natural characteristics or human activities, which may not optimally align with the discriminative features and class relationships present in remotely sensed imagery. This paper explores a novel cluster-based class hierarchy modeling framework that generates data-driven hierarchical structures for LULC semantic segmentation. First, we perform spectral clustering on confusion matrices generated by a flat model, and then we introduce a hierarchical cluster validity index to obtain the optimal number of clusters to generate initial class hierarchies. We further employ ensemble clustering techniques to yield a refined final class hierarchy. Finally, we conduct comparative experiments on three benchmark datasets. Results demonstrating that the proposed method outperforms predefined hierarchies in both hierarchical LULC segmentation and classification.