Evaluating Urban Perception: Using Explainable Machine Learning Predict Through the Best Pipeline

Understanding subjective urban experiences is essential for designing cities that enhance well-being. Urban design should account for the psychological effects of environments on individuals, as these significantly shape perceptions and behaviors. However, a major challenge is the limited availability of urban perception data. Recent studies have leveraged large, crowdsourced datasets like Place Pulse 2.0 (PP2) to inform machine learning (ML) models for urban perception prediction, but the accuracy and reliability of outcomes remain underexplored. There is a critical need to evaluate whether these datasets truly capture human perceptions. This study investigates the role of urban street images in understanding environmental perceptions, using the PP2 dataset and ML techniques. It explores various ML pipelines, employing TPot AutoML for model selection and 5-fold cross-validation to prevent overfitting. The goal is to identify the most efficient model that strengthens the link between automated predictions and human perception. The study also applies SHAP (SHapley Additive exPlanations) to interpret model outputs, revealing feature importance and interactions. This improves transparency and ensures ML-generated insights are actionable for urban planning. By rigorously testing ML pipelines, this research enhances predictive accuracy and contributes to the development of reliable urban design tools. The findings highlight ML’s potential in processing large-scale perception data, uncovering hidden patterns, and informing people-centered urban planning. However, further validation against real-world surveys is necessary to ensure robustness and generalizability in assessing urban perceptions.