From satellite imagery to perceived noise annoyance: estimating urban sound environment annoyance using remote sensing and machine learning

High-quality acoustic management constitutes a core element in promoting the long-term well-being and growth of urban settings. Evaluating the sound environment is a critical step in assessing urban liveability. Existing evaluation approaches primarily rely on objective measurements such as sound pressure level (SPL) and the creation of urban noise maps. However, SPL-based metrics do not fully capture residents’ subjective responses to acoustic environments. To address these limitations, this study proposes a novel framework that integrates subjective perception data with remote sensing for urban sound environment. Specifically, residents’ perceived annoyance towards environmental noise is defined as ‘sound environment annoyance (SEA)’ following established soundscape frameworks, and used as a proxy for subjective sound perception. SEA is then estimated using machine learning and multi-scale, multispectral remote sensing imagery. A questionnaire survey was conducted in the urban core to collect residents’ SEA (n = 1190) using a five-point Likert scale from very comfortable (1) to very uncomfortable (5). Trained investigators guided respondents to ensure standardized completion. The survey and Sentinel-2 imagery acquisition were both conducted in 2024 to ensure temporal consistency between perceived SEA and environmental observations. Based on Sentinel-2 imagery, we extracted remote-sensing features capturing multiple spatial scales, spectral bands, and neighbourhood contexts. An XGBoost model was trained to link these features with SEA, and model predictions were evaluated using standard performance metrics. The model achieved an MAE of 0.32, an RMSE of 0.48, and an R2 of 0.65 on the test set, demonstrating strong predictive performance. Furthermore, we applied the SHAP method to interpret the model and identify key remote sensing features that contribute to the predictions. The proposed framework enables rapid, low-cost, and spatially continuous estimation of urban sound environment perception. This work contributes to cross-modal research in environmental science and offers a novel pathway for understanding urban sound environment.