Author(s): Seunghyun Hwang; Jongyun Byun; Seoyeong Ku; Jongjin Baik; Changhyun Jun
Linked Author(s):
Keywords: Convolutional neural networks; IoT-based rainfall observation; Rainfall acoustics; Raindrop size distribution; Spectrogram analysis
Abstract: The raindrop size distribution (DSD) regulates fundamental hydrometeorological properties, such as precipitation type, intensity, and terminal velocity, and provides the physical basis for precipitation retrieval in all radar- and satellite-based remote sensing algorithms. However, its observation has predominantly relied on expensive instrumentation, thereby limiting the establishment of dense measurement networks capable of resolving spatial heterogeneity. To address this issue, this study presents a cost-efficient acoustic sensing approach for estimating DSD using a lightweight IoT-based platform and artificial intelligence. Rainfall-induced sound was collected under natural conditions using a Raspberry Pi–based acoustic sensor, which was co-located with an optical disdrometer to ensure synchronized reference observations. The recorded acoustic signal was divided into 10-second segments, each of which was converted to a spectrogram using short-time Fourier transform. These spectrograms were then supplied to convolutional neural networks trained to predict number concentration per diameter interval across 30 Parsivel2-defined drop-size classes. Although a limited number of weak rainfall signals were falsely detected during no-rain intervals, the model effectively reconstructed the observed DSD structure with consistent accuracy. This result indicates that acoustic signals generated by raindrop impacts retain sufficient microphysical information to enable class-wise DSD inference at fine temporal resolution.
Year: 2026