Feng et al. (2025) A novel deep learning approach for high-precision rainfall intensity inversion using urban surveillance audio

Identification

• Journal: Advances in Space Research
• Year: 2025
• Date: 2025-10-24
• Authors: Jiangfan Feng, Xi Fu, Shaokang Dong
• DOI: 10.1016/j.asr.2025.10.070

Research Groups

• School of Artificial Intelligence, Chongqing University of Posts and Telecommunications, Chongqing, China
• Key Laboratory of Tourism Multisource Data Perception and Decision, Ministry of Culture and Tourism (TMDPD, MCT), Chongqing University of Posts and Telecommunications, Chongqing, China
• School of Computer Science and Technology, Chongqing University of Posts and Telecommunications, Chongqing, China

Short Summary

This paper introduces MS-TF RainNet, a novel deep learning framework for high-precision rainfall intensity inversion using urban surveillance audio, achieving an RMSE of 0.7708 mm/h and outperforming a Transformer-based baseline by 14.94% in RMSE under denoised conditions.

Objective

• To develop a novel deep learning framework (MS-TF RainNet) that enables high-precision rainfall intensity inversion from urban surveillance audio, addressing limitations in current audio-based approaches regarding multi-scale pattern capture, time–frequency dependency modeling, and effective regression constraints.

Study Configuration

• Spatial Scale: Urban environments, focusing on leveraging existing urban surveillance infrastructure.
• Temporal Scale: Continuous, fine-grained monitoring of rainfall intensity.

Methodology and Data

• Models used: MS-TF RainNet, a deep learning framework comprising:
  • A hierarchical multi-scale feature extraction module processing Mel Frequency Cepstral Coefficients (MFCC) through parallel convolutional branches.
  • A dual-domain attention mechanism combining temporal and frequency attention.
  • Deep supervision with auxiliary regression heads.
• Data sources: Surveillance Audio Rainfall Intensity Dataset (SARID).

Main Results

• MS-TF RainNet achieved a Root Mean Square Error (RMSE) of 0.7708 mm/h and an R² of 0.8196 under denoised conditions.
• It outperformed a Transformer-based baseline model by 14.94% in RMSE and 9.18% in R².
• In noisy environments, the model maintained robustness with an RMSE of 0.8443 mm/h and an R² of 0.7983.

Contributions

• Proposes MS-TF RainNet, a novel deep learning framework for high-precision rainfall intensity inversion from urban surveillance audio.
• Introduces a hierarchical multi-scale feature extraction module to capture both local and global rainfall patterns.
• Incorporates a dual-domain attention mechanism for transient noise suppression and spectral feature amplification.
• Utilizes deep supervision with auxiliary regression heads to enforce hierarchical feature consistency and mitigate gradient vanishing.
• Offers a cost-effective and transformative solution for urban hydrometeorology by leveraging existing surveillance infrastructure, outperforming conventional methodologies.

Funding

Not explicitly mentioned in the provided text.

Citation

@article{Feng2025novel,
  author = {Feng, Jiangfan and Fu, Xi and Dong, Shaokang},
  title = {A novel deep learning approach for high-precision rainfall intensity inversion using urban surveillance audio},
  journal = {Advances in Space Research},
  year = {2025},
  doi = {10.1016/j.asr.2025.10.070},
  url = {https://doi.org/10.1016/j.asr.2025.10.070}
}