The detection of rain on snow events as well as precipitation phase partitioning is becoming increasingly important in the wake of climate change, the elevated need for more accurate hydro logic models, and the increase of rain on snow related flood events. The project titled “Advancing Snow Observation Systems to Improve Operational Streamflow Prediction Capabilities” at the University of Vermont focuses on creating a low-cost, low-power network of embedded devices, which will run a machine learning model to perform precipitation phase partitioning as well as rain on snow event detection and provide near real-time (NRT) data reporting. This thesis outlines how we are training this machine learning model to differentiate precipitation phases from audio data, also known as acoustic disdrometry. We have utilized both simulated and real precipitation data, recorded using a low cost embedded device developed at UVM. Various audio features and machine learning models have been investigated to determine how to achieve the highest possible accuracy in weather classification from audio files. On simulated data, we have achieved up to 98.3% accuracy and on real data we have achieved up to 91.58 % accuracy. These results provide a promising proof of concept for future system deployments.
Rachael Lauren Chertok (Wed,) studied this question.