Accuracy Metrics for Explainable Unsupervised Methods in Remote Sensing: A Scorecard for Validity and Explanation Quality

Kurzfassung

Unsupervised learning is widely used in remote sensing when labels are sparse, delayed, or heterogeneous across regions and seasons. Its operational use is limited when cluster outputs are not accompanied by transparent quantitative reporting. We propose a compact scorecard for explainable unsupervised pipelines that separates cluster validity, explanation quality, and seed-based reproducibility in no-label settings. We focus on explainable k-means (XKMeans) and explainable Gaussian mixture models (XGMM) and link each scorecard quantity to concrete explanation artifacts: model-tied cluster feature sets and surrogate-based summaries. Validity is quantified through silhouette and, when benchmark labels exist, purity and normalized mutual information. Explanation quality is evaluated through fidelity, coverage, sparsity, and stability. We demonstrate the scorecard on a Sentinel-2 wildfire scene without pixel-level ground truth. The experiment is intentionally a compact proof-of-concept in a low-dimensional feature space, while broader validation on higher-dimensional and multimodal remote-sensing settings is a natural next step.