Embedded IoT System for Acoustic Precipitation Phase Partitioning via Edge ML and MFCCs

Accurate and real-time detection of precipitation phases is essential for hydrological modeling, water resource management, and climate impact assessments. However, conventional methods struggle to distinguish precipitation phase near freezing and are unsuitable for distributed deployment in remote, complex terrain due to cost, size, and power constraints. In this work, we present an integrated acoustic sensing system that combines machine learning on the edge with acoustic sensing and Mel-Frequency Cepstral Coefficients (MFCC)-based feature extraction, all implemented on our designed edge device. The device is low-cost, with LAN and WAN capabilities for near real-time reporting, and can integrate multiple sensors for snow hydrology studies, in addition to detecting the precipitation phase. We present the design, fabrication, and validation of the Advanced Unified Rainfall and Atmospheric Monitoring (AURA) system, integrating sensors for temperature, humidity, ultrasonic snow depth, solar radiation, and wind velocity. AURA features an embedded machine learning algorithm that enables accurate classification of precipitation phases. Short-time Fourier Transform (STFT) and MFCCs are computed on the edge from recorded precipitation acoustics via novel Mel filter bank computation methods. Using support vector machine (SVM) and random forest (RF) classifiers, the RF model achieves testing accuracy of 97.35% on simulated precipitation acoustics and 85.67% in consolidated class (CC) environmental recordings. The SVM classifier achieves 98.07% accuracy on simulated acoustics and 85.99% on CC environmental recordings. The low-power AURA network uses a LoRa star topology with TDMA and an Iridium gateway for reliable data transfer from remote sites. SNR analysis and comprehensive field tests confirm and validate system performance.