Examining the complex and cumulative effects of environmental exposures on noise perception through interpretable spatio-temporal graph convolutional networks.

With the continuous expansion of urban populations, interactions between humans and their environments have grown increasingly frequent and complex. Traditional environmental studies often focus on isolated factors or specific locations, neglecting the cumulative impact of contextual environments on human perception. To address this gap, this study employs noise exposure as a case study and utilizes an interpretable spatio-temporal graph convolutional network (ST-GCN) framework to model the perception process in urban environments. The results indicate that the ST-GCN model outperforms other models in predicting noise perception, 10 min of short-term exposure represents a critical threshold for detecting environmental noise, whereas a 15-minute duration serves as the perceptual threshold for predicting noise annoyance. While sound pressure levels primarily drive perception, other environmental features significantly contribute, with annoyance being more modulable. Air quality and sound pressure levels exhibit a synergistic interaction, amplifying both objective noise perception and annoyance. Conversely, green space mitigates these perceptions, while walls and buildings play a dual role: reducing perceived noise in the objective noise model but enhancing annoyance in the noise annoyance model. These findings reveal the complex spatio-temporal dynamics of environmental exposure on noise perception, providing theoretical insights for urban environmental management and resident well-being research.