Explainable unsupervised Methods for Forest Fire Detection

Abstract

Wildfires are among the most damaging natural hazards, and reliable satellite-based detection is essential for operational response. Many current pipelines for multispectral Sentinel-2 imagery are supervised and data-hungry, and they often behave as black boxes, limiting trust and expert auditability during fast-evolving events. We present a fully unsupervised and intrinsically explainable framework for wildfire detection that combines explainable K-Means (xKMeans), explainable Gaussian Mixture Models (xGMM), and explainable Latent Dirichlet Allocation (LDA). Each pixel is represented by a compact fire sensitive feature vector built from Sentinel-2 bands B08, B8A, and B12 together with the Normalized Burn Ratio (NBR) and the Mid-Infrared Burn Index (BAIS1). The three models expose complementary structure: xKMeans yields human-readable, rule based partitions; xGMM provides probabilistic memberships and entropy-based uncertainty maps that highlight ambiguous transition zones; and LDA offers global, topic-level explanations by summarizing co-occurring spectral-index patterns into interpretable fire regimes. We demonstrate the approach on multiple wildfire scenes, including case studies in Los Angeles (USA) [1], Greece [2], and an additional site in Romania based on GDACS reported event metadata. We report qualitative and label-light quantitative consistency analyses based on model agreement and perimeter-aligned evaluation. The results indicate that the proposed workflow delineates fire-affected areas while remaining transparent enough for expert scrutiny, offering a label-efficient alternative to supervised deep-learning systems for operational wildfire monitoring.