This paper introduces a framework for predicting grapevine cold hardiness—the temperature at which 50% of dormant buds freeze (LT50)—across different geographic regions. By learning transferable latent representations, the authors aim to provide accurate, actionable predictions for growers in data-scarce regions where local historical data is insufficient for traditional modeling.
The Challenge of Regional Variability
Grapevine cold hardiness is critical for managing frost risks, but measuring it requires specialized lab equipment and destructive sampling, making real-time monitoring infeasible. While existing biophysical, hybrid, and deep learning models perform well on local data, they are site-specific and struggle to generalize to new regions. The authors note that the same cultivar often exhibits different cold hardiness responses across regions due to unobserved factors like soil properties, management practices, and local climate.
A Unified Modeling Framework
The researchers propose a multi-task learning framework that treats cold hardiness prediction as a sequence prediction problem. The model consists of three components:
Task Encoder ($\phi$): Maps specific combinations of regions, weather stations, vineyards, and cultivars into a latent feature space.
Temporal Encoder ($\psi$): Encodes the relationship between weather history and crop state based on the latent task representation.
Biophysical Model ($\mathcal{M}$): Uses the output of the previous components to produce an LT50 prediction.
This structure allows the framework to unify biophysical, deep learning, and hybrid modeling paradigms, enabling a systematic comparison of their performance.
Enabling Transfer to New Regions
To support predictions in regions or for cultivars where no prior data exists, the authors developed two methods to infer the latent task representation ($\phi$):
Text-based Transfer: Uses an E5 sentence embedding model to encode descriptive text about a cultivar or region (e.g., climate, soil, and hardiness characteristics). This supports zero-shot transfer to previously unseen locations.
Data-based Transfer: Uses a cross-attention mechanism to process limited, irregular historical LT50 observations and weather data. This supports few-shot transfer when some local data is available.
Performance and Evaluation
The authors evaluated their approach using datasets from six North American regions: British Columbia, Ontario, Washington, New York, Michigan, and Nova Scotia. Experiments showed that their approach consistently outperformed state-of-the-art methods in predictive accuracy. The researchers observed that latent representations clustered primarily by region rather than by cultivar, confirming that regional characteristics exert a stronger influence on cold hardiness than cultivar identity alone. The model has been deployed on AgWeatherNet for the 2026-2027 dormancy season.
Comments (0)
to join the discussion
No comments yet
Be the first to share your thoughts!