Explainability of Deep Learning models for Urban Space perception

Explainability tools can reveal hidden insights for urban planning: By using GradCAM visualization techniques on deep learning models, researchers were able to identify which landscape objects influence AI predictions about urban space quality (such as safety or beauty), helping urban planners understand what factors the models actually consider important.

Current object detection methods may be incomplete: The study discovered new objects relevant to urban perception that were not represented in existing object detection models previously used for annotation, suggesting that AI-assisted urban analysis tools may be missing important environmental features that affect how people perceive their surroundings.

Model architecture choice matters for explainability: The research found that transformer-based architectures work better with GradCAM explainability techniques compared to traditional Convolutional Neural Networks, indicating that when deploying AI for urban planning decisions, the choice of model architecture affects how well planners can understand and trust the AI's reasoning.

Explainability is essential before deployment in policy contexts: The paper demonstrates that without explainability techniques, urban planners using "blackbox" deep learning models would lack transparency into what the AI is actually measuring, which is critical when these tools inform real-world decisions about urban environments that affect residents' quality of life.