A deep learning pipeline for rainfall estimation from surveillance audio

Manuel Fiallos-Salguero, Soon-Thiam Khu, Mingna Wang,
A deep learning pipeline for rainfall estimation from surveillance audio,
Journal of Hydrology,
Volume 662, Part A,
2025,
133921,
ISSN 0022-1694,
https://doi.org/10.1016/j.jhydrol.2025.133921.
(https://www.sciencedirect.com/science/article/pii/S0022169425012594)
Abstract: Rainfall plays a crucial role in hydrological studies and modeling; however, existing measurement methods often face notable errors and limitations. To address these challenges, this study explores the use of widely available devices, such as surveillance cameras, as a cost-effective alternative to enhance the accuracy and spatiotemporal data coverage in urban environments. This study introduces a novel audio-based deep-learning pipeline to enhance data quality and estimate rainfall intensity from environmental audio captured by surveillance cameras. The proposed framework integrates two lightweight deep networks to handle complex data patterns effectively. The first network isolates and restores rainfall signals by suppressing overlapping ambient noise, while the second estimates rainfall intensity from the enhanced acoustic signals. Validation on real-world rainfall events demonstrated the effectiveness of our method in noise reduction, achieving a mean signal-to-noise ratio of 7.28 dB, a spectral subtraction ratio of 6.58 dB, a mean absolute error of 0.038, and a root mean square error of 0.063. With the improved signal quality, our predictive model achieved high accuracy, with R2 scores ranging from 0.83 to 0.88 compared to rain gauge data. Besides, cumulative rainfall estimations from surveillance audio yielded a mean absolute percentage error ranging from 6.05 to 18.04 %, revealing significant improvements over existing methods. Beyond accuracy, the proposed approach employs robust and lightweight model architectures, especially in the denoising stage, enabling efficient deployment with minimal computational cost to existing surveillance infrastructure. This study highlights the potential of audio-based rainfall estimation for real-time monitoring, offering a scalable and cost-effective solution for urban hydrological applications and disaster response systems.
Keywords: Rainfall estimation; Surveillance audio; Deep learning; Spatial-temporal resolution; Signal processing; Noise reduction