Role of Visualization in Explainable AI : diverse EO case studies

Abstract

Overview In XAI, visualization simplifies complex information about "black-box" AI models into accessible formats, helping users understand model behavior, evaluate predictions, and gain trust in the AI system. By creating visual representations of features, decision-making processes, and data relationships, visualization techniques provide a fine-grained perspective of the AI's internal workings, enabling users to identify patterns, debug models, and compare different approaches for better decision-making . In this poster, we convert satellite radar images into stable colour-coded maps using BoVW→LDA textualisation and elastic search over words/topics. We provide three transparent interactions—content similarity, content search by word/topic, and sub-content match—tightly coupled with side-by-side original vs. map views. Each area carries an explicit confidence bar, last-update time, and a short history strip to communicate reliability. Known limits (rough seas, rapid melt/freeze edges, sub-tile objects) are shown through lower confidence. A lightweight feedback button records uncertain spots for review and iterative improvement. Highlights • Color-coded, georeferenced maps derived from radar image patterns • Three interactions: similarity, term/topic search, visual-phrase match • Side-by-side original quick-look and map for visual verification • Confidence, last-update, and recent history displayed on-map • Clear communication of limits; confidence reflects difficult conditions • Simple feedback channel for continuous refinement Projects • H2020 ExtremeEarth • HGF AutoCoast • AI4EU