D4.4 Report on the methodology of transparent ML models (full version)

This report presents a methodology for building transparent and trustworthy machine learning (ML) models tailored to remote digital tower (RDT) applications. Current object detection models like YOLO and R-CNN often lack robustness, domain adaptability, and interpretability, which makes them difficult to trust in critical systems. To address these challenges, this work develops a hybrid framework combining interpretable deep learning techniques, fairness-aware data strategies, and human-in-the-loop design. For video data, methods like GradCAM and YOLO-World are used to visualize and explain object detections, while audio data is interpreted using LIME and SHAP to clarify decisions made from pilot-controller communication. Active learning reduces training data needs, and fairness is improved using SMOTE-driven oversampling methods. The framework is grounded in human-centred explainable AI (XAI), integrating insights from workshops with air traffic controllers to ensure meaningful, practical explanations. This approach supports more accountable, fair, and transparent AI systems ready for real-world deployment in air traffic management.