Accurate winter precipitation-type (p-type) forecasts are critical for public safety and effective decision-making during hazardous weather events. This project aims to improve the accuracy, quantify uncertainty, and ensure the physical consistency of winter weather p-type analyses and forecasts. We are developing an advanced AI-based p-type system, along with tailored visualizations and documentation, to deliver actionable information for decision-makers across various forecast lead times.

Key achievements include the development of an evidential p-type model that provides both high accuracy and robust uncertainty estimates. We have successfully applied this model to High-Resolution Rapid Refresh (HRRR) hindcasts. Additionally, we are refining an interview protocol in collaboration with our Risk Communication team and building a visualization website to enhance communication of p-type hindcasts. The project also supports workforce development through mentorship of multiple interns, including three SIParCS interns (two in 2023, one in 2024) and a Vaisala intern in 2023.

Related Project: Trustworthy AI for Winter Precipitation Type with Users

Accurately detecting precipitation in real-time can assist in state-wide transportation and operations, especially during cold and icy winter conditions. This project aims to develop an AI-powered system that uses explainable and trustworthy methods to detect the presence or absence of precipitation streaks in New York State Mesonet (NYSM) camera images during both daytime and nighttime. Integrated with the NYSM web-based interface, this automated detection system will provide real-time insights for weather monitoring. Our Winter Weather team is collaborating with the Albany Weather Forecast Office (WFO) to align with operational needs, and establishing robust labeling methods for training machine learning models.

Related Project: Leveraging Social Science for Reproducible Supervised ML Hand Labeling

Numerical Weather Prediction (NWP) models play a vital role in forecasting critical weather variables such as temperature, wind, and precipitation. However, these models often exhibit errors and biases due to varying physical processes, spatial resolutions, and regional characteristics. This project aims to quantify and predict NWP errors across the New York State (NYS) and Oklahoma (OK) Mesonets using machine learning (ML) approaches. Our goal is to develop accurate, region-specific ML models that forecast temperature, wind, and precipitation errors, providing user-tailored insights for National Weather Service (NWS) partners.

Our team recentl publishing a peer-reviewed study verifying major NWP models using NYS Mesonet data and developing a Long Short-Term Memory (LSTM) model that successfully predicts forecast errors. To overcome challenges like model variability and regional differences, we are conducting an extensive ML architecture intercomparison and analyzing the relationship between NWP errors and geographic factors (e.g., land-use, topography). We are strengthening partnerships with NWS offices to co-develop practical, user-focused deliverables, ensuring that our ML-driven insights enhance operational weather forecasting and decision-making.