Fiallos-Salguero et al. (2025) A deep learning pipeline for rainfall estimation from surveillance audio

Identification

• Journal: Journal of Hydrology
• Year: 2025
• Date: 2025-07-16
• Authors: Manuel Sebastian Fiallos-Salguero, Soon‐Thiam Khu, Mingna Wang
• DOI: 10.1016/j.jhydrol.2025.133921

Research Groups

• School of Environmental Science & Engineering, Tianjin University, China
• School of Civil Engineering, Tianjin University, China

Short Summary

The study develops a deep learning pipeline that utilizes audio captured by surveillance cameras to estimate rainfall intensity, employing a two-stage network for noise suppression and intensity prediction.

Objective

• To create a cost-effective, scalable, and high-spatiotemporal resolution alternative to traditional rain gauges and radar for urban rainfall monitoring.

Study Configuration

• Spatial Scale: Urban environments (local scale).
• Temporal Scale: Real-time monitoring (targeting high-resolution requirements of 1–5 min).

Methodology and Data

• Models used: A pipeline consisting of two lightweight deep learning networks: one for signal denoising (isolating rainfall signals from ambient noise) and one for rainfall intensity estimation.
• Data sources: Environmental audio recorded via surveillance cameras and ground-truth data from rain gauges.

Main Results

• Denoising Performance: Achieved a mean signal-to-noise ratio (SNR) of 7.28 dB, a spectral subtraction ratio of 6.58 dB, a mean absolute error (MAE) of 0.038, and a root mean square error (RMSE) of 0.063.
• Intensity Estimation: The predictive model achieved $R^2$ scores between 0.83 and 0.88 compared to rain gauge measurements.
• Cumulative Rainfall: The mean absolute percentage error (MAPE) for cumulative estimations ranged from 6.05% to 18.04%.

Contributions

• Proposes a novel, lightweight audio-based framework that leverages existing urban surveillance infrastructure, reducing the need for expensive dedicated sensor networks while improving spatiotemporal data coverage for urban hydrological applications and disaster response.

Funding

• Not specified in the provided text.

Citation

@article{FiallosSalguero2025deep,
  author = {Fiallos-Salguero, Manuel Sebastian and Khu, Soon‐Thiam and Wang, Mingna},
  title = {A deep learning pipeline for rainfall estimation from surveillance audio},
  journal = {Journal of Hydrology},
  year = {2025},
  doi = {10.1016/j.jhydrol.2025.133921},
  url = {https://doi.org/10.1016/j.jhydrol.2025.133921}
}