Rainfall intensity estimation via raindrop sounds leveraging convolutional neural networks and low-cost IoT sensors

This study proposes a convolutional neural networks (CNNs)-based framework for estimating rainfall intensity using acoustic signals acquired from raindrops. Raindrop sounds were collected under real-world conditions using an internet of things (IoT) sensor-based acoustic data collection device. The collected signals were then transformed into spectrotemporal representations via short-time Fourier transform (STFT) and mel-spectrogram analysis. A dual-stream CNNs model was constructed to learn from both spectrogram types, leveraging their complementary strengths in capturing high- and low frequency signal characteristics across various rainfall intensities. The model was trained using a balanced dataset representing no rain, weak, moderate, and heavy rainfall, and validated against ground truth measurements from an optical disdrometer (i.e., OTT PARSIVEL²). Evaluation results indicate that the proposed method yields promising performance, with a root mean square error of 4.89 mm/h, a mean absolute error of 2.02 mm/h, and a R² of 0.75. While the model effectively estimates weak to moderate rainfall, it tends to underestimate extreme rainfall events due to their underrepresentation in the training data. These findings demonstrate the feasibility of rainfall intensity estimation from acoustic signals and highlight the potential of deep learning-based acoustic sensing for hydrometeorological applications in observation-challenged areas.