NOVEMBER 2025
Forecasting Frost Before It Hits
Frost is one of the most damaging weather risks in agriculture, and growers often have to make protection decisions with limited time and imperfect information.
The Frost Risk Data Challenge explored how machine learning, spatial modeling, and high-performance computing could improve early frost detection and support faster, more informed action in the field.
OVERVIEW
A data challenge built around an operational problem
The Frost Risk Data Challenge invited university teams to develop predictive models using localized microclimate weather data from California. The goal was not just better model performance, but more useful forecasting for agricultural decision-making, with methods designed to be transferable to other regions and cropping systems.
Supported by the San Diego Supercomputing Center and the National Data Platform, the challenge gave participants access to real-world datasets and supercomputing resources to test approaches grounded in both technical rigor and agricultural relevance. Entries were evaluated on accuracy, creativity, and overall solution effectiveness.
CHALLENGE STRUCTURE
Interdisciplinary teams, applied modeling, and reproducible results
Launched in November 2025, the challenge brought together undergraduates, graduate students, faculty, staff, and independent contributors to work on frost risk forecasting through data-driven methods. Teams developed machine learning and spatial models aimed at delivering earlier, more reliable frost predictions for growers and agricultural operators.
The challenge emphasized reproducible work and practical problem-solving. Participants were asked to build models that could perform under real conditions, not just in theory, and to translate technical approaches into tools that could support proactive decisions in agricultural operations.
WINNING TEAMS
All teams that submitted valid and reproducible entries received a Certificate of Participation. The top three teams earned an official F3i Digital Credential through an open badging system for professional and academic use. Leading teams were also invited to present their work to F3i’s grower and industry partners and may be considered for future sprint design teams and commercialization cohorts.
Recognition and rewards
Cash prizes were awarded to the top three teams: $1,500 for first place, $750 for second, and $400 for third.
1st Place — FrostByte
San Jose State University
Probabilistic forecasting system built on XGBoost, with spatial validation and an interactive decision-support dashboard.
2nd Place — TylerAziz
University of California, Davis + California Polytechnic State University, San Luis Obispo
Physics-informed tree-based machine learning using Random Forest and XGBoost with strict station-grouped cross-validation for spatial generalization.
3rd Place — AgriFrost AI
Independent Contributor
Systematic feature engineering with class-balanced gradient boosting using LightGBM for probabilistic frost risk forecasting.