Acoustic Rainfall Sensing A Data-Driven Approach For Urban Rainfall Intensity Estimation

Abstract

Accurate rainfall estimation in urban environments remains a significant challenge due to the limitations of traditional rain gauge networks and weather radar calibration issues. This study explores the potential of acoustic rainfall sensing, leveraging audio data recorded from rainfall impacting various urban surfaces. The rainfall audio data collection was conducted in Monash University Malaysia campus over two years, using professional recorders at five different locations. A data-driven approach was employed using artificial neural networks (ANN) and extreme gradient boosting (XGBoost) models to develop an acoustic rainfall estimation model. Results demonstrated that a combination of loudness, frequency, and cepstral domain features significantly improved prediction accuracy, with the ANN model outperforming XGBoost. The ANN model demonstrated consistent performance across the training (R² = 0.675, RMSE = 0.287 mm/min, MAE = 0.203 mm/min) and validation datasets (R² = 0.681, RMSE = 0.286 mm/min, MAE = 0.203 mm/min). The findings suggest that acoustic sensing, when integrated with urban IoT frameworks, can serve as a low-cost and scalable alternative for urban rainfall monitoring. Future research should focus on enhancing feature selection techniques and expanding real-world testing environments to improve model robustness.