Explainable clustering in remote sensing

Kurzfassung

Understanding how unsupervised clustering methods form their decision boundaries remains a major challenge in remote sensing explainability. Although algorithms such as k-means are computationally efficient and widely used, their cluster assignments are difficult to interpret and provide limited insight into the underlying physical or semantic structure of the data. Existing explainable approaches for unsupervised learning offer partial transparency, but often lack rule-based semantic explanations and provide limited support for incorporating domain knowledge in an interpretable manner. This thesis aims to contribute to this research direction by adding explainability to clustering frameworks that combines k-means clustering and Latent Dirichlet Allocation (LDA) by replacing k-means with X-kMeans variants to enhance interpretability while quantifying divergence from the original k-means assignments and confidence in the resulting explanations. Cluster labels produced by k-means are approximated using rule-based decision trees, enabling explicit and human-readable descriptions of cluster boundaries. Both greedy and non-greedy Iterative Mistake Minimization strategies are investigated to analyze trade-offs between model compactness, fidelity, and semantic richness, as well as the transferability of learned rules to unseen data. LDA is subsequently applied to derive semantic topics from cluster-aligned feature representations. Beyond structural explainability, this work introduces two novel metrics—the Topic Alignment Factor (TAF) and the Word Explainability Confidence Score (WECS)—to quantitatively assess the reliability and semantic consistency of topic–word associations. In addition, Cluster Fidelity is used to measure how faithfully X-kMeans reproduces k-means cluster assignments. The proposed framework is evaluated on subsets of the UCMerced remote sensing dataset. Experimental results indicate that non-greedy decision trees achieve improved semantic alignment and topic coherence, while greedy trees retain compactness and comparable fidelity to k-means. Overall, this work contributes a confidence-aware structural–semantic explainable clustering framework that improves transparency, trustworthiness, and interpretability of unsupervised learning in remote sensing applications