AI for Remote Sensing and Sustainable Cities - UGOD Thrust Seminiar | Topic 1: Concept-based explanation methods for deep learning with optical remote sensing; Topic 2: Artificial Intelligence for Traffic Safety between Vehicles and Vulnerable Road Users

Topic 1: Concept-based explanation methods for deep learning with optical remote sensing

While deep learning models have seen widespread adoption for a wide range of remote sensing prediction tasks, it remains difficult to understand their prediction patterns and to extract new knowledge from them. Meanwhile, emerging legal frameworks have increasingly pushed for a right to an explanation for predictions made by AI models. Although the need is clear, recent studies have shown that RS-based studies are especially lacking in explanation methods, as research approaches that are devised for natural images often need adaptation before they can be used in a remote sensing setting. One such approach is the use of concept-based explanation methods, which use intermediate concepts to help explain a target variable. Research has demonstrated that bottleneck models can deliver performance while providing explanations that are understandable for humans. In this presentation, I will introduce concept-based explanation methods for deep learning, with a specific focus on remote sensing applications. In particular, I will discuss semantic bottlenecks and our recent work on testing with concept activation vectors for remote sensing applications.

Topic 2: Artificial Intelligence for Traffic Safety between Vehicles and Vulnerable Road Users

Traffic safety stands as the cornerstone of vehicular environments, especially with the advent of artificial intelligence (AI)-based systems like self-driving cars. Critical areas within urban settings, such as intersections and shared spaces, pose heightened risks where vehicles and vulnerable road users (VRUs) like pedestrians and cyclists directly interact. The VeVuSafety project endeavors to enhance safety between vehicles and VRUs by advancing state-of-the-art artificial intelligence methodologies. This project aims to construct deep learning frameworks capable of comprehending road users' behavior across diverse mixed traffic scenarios. Leveraging heterogeneous traffic data from sources such as cameras and laser sensors, our approach focuses on multimodal behavior learning of various road user types, encompassing detection, tracking, and motion prediction, alongside interpreting the driving environment.