Audio-Based Rainfall Classification Using Spectrograms and YOLOv8n on Embedded Systems

Hydrogeological risks such as floods and landslides increasingly threaten critical infrastructure, emphasizing the need for continuous and reliable rainfall monitoring. Traditional sensing approaches often face limitations in spatial resolution, latency, and installation cost. This paper presents an audio-based rainfall classification system using spectrogram analysis combined with a lightweight YOLOv8n neural network, optimized for embedded deployment. A custom acquisition kit was developed to capture multimodal environmental data, although this study focuses exclusively on the audio channel. Audio signals sampled at 8 kHz were segmented, labeled into six rainfall classes through an automated and manually verified procedure, and converted into spectrograms. Training was performed on a dedicated dataset collected over two rainy days. Experimental results demonstrate a weighted accuracy of 86.99% and a weighted F1-score of 86.89% on unseen data, confirming the potential of acoustic sensing for rainfall classification. Future extensions will integrate multimodal data fusion to enhance robustness in noisy environments and further improve system performance.