In brief
Urban acoustic pollution exhibits strong cyclic temporal patterns driven by human activity alongside spatial correlations governed by street topology and urban morphology. Integrating 1D temporal convolutions, recurrent LSTM memory, and graph transformer layers (CNN1D+LSTM+TransformerConv) delivers highly accurate predictive mapping across Madrid's urban sensor grid.
Contribution
A hybrid CNN1D, LSTM and graph-transformer model evaluated against temporal, convolutional and graph-based alternatives.
Key finding
Explicit graph structure produced the best prediction accuracy, reaching an RMSE of 0.0169 and a correlation coefficient of 0.9601.
Models and data
Distinctive research elements
Models and methods
- CNN1D+LSTM+TransformerConv
- Transformer
Data and evaluation
- Madrid urban-noise observations
Author-written overview
Research overview
Urban noise varies across both time and the structure of a city. Using Madrid as a case study, this research compares convolutional, recurrent and graph-based deep-learning approaches for noise prediction. The models test complementary ways of representing local patterns, longer temporal dependencies and relationships between observation locations. The hybrid CNN1D+LSTM+TransformerConv architecture produces the strongest result, with an RMSE of 0.0169 and a correlation coefficient of 0.9601. Its error is 5.1% lower than that of the second-best evaluated model. The comparison provides evidence that an explicit graph representation adds useful spatial information to temporal sequence modelling for urban-noise forecasting.
Cite this work
BibTeX
@article{semper2025noise,
title = {Noise Pollution Prediction in a Densely Populated City Using a Spatio-Temporal Deep Learning Approach},
author = {Marc Semper and Manuel Curado and Jose Luis Oliver and Jose F. Vicent},
journal = {Applied Sciences},
year = {2025},
volume = {15},
number = {10},
eid = {5576},
doi = {10.3390/app15105576}
}Always check the publisher record for final volume, issue and page information before citing.