Multi-Task Learning for Sparsely-Labeled Time Series: A Case Study on Cold-Hardiness Modeling

Predicting grape cold hardiness (the temperature at which lethal freezing occurs) and budbreak is critical for agricultural frost mitigation. However, these processes are difficult to measure, and ground-truth data is often temporally sparse and limited for specific plant cultivars. This paper investigates whether multi-task learning (MTL) can leverage shared biological structures across different grape cultivars to improve prediction accuracy for individual cultivars, especially those with limited data.

The authors employ recurrent neural networks (RNNs) as a base model to process daily weather time series. They implement two "wrapper-based" MTL architectures that can be easily applied to existing single-task models:

1. Multi-Head Models: A shared backbone network feeds into task-specific output heads for each cultivar.
2. Task Embedding Models: The input is augmented with a learned dense embedding vector representing each cultivar, allowing the model to capture task-specific dynamics.

These models were evaluated against single-task learning (STL) and the current state-of-the-art scientific model (the Ferguson model) across 21 grape cultivars. The authors also explored transfer learning settings where models are adapted to new cultivars without access to the original source data.

MTL consistently outperformed both STL and the state-of-the-art scientific models. A key finding is that MTL models trained on only 2 to 5 seasons of data for a cultivar could match or exceed the performance of STL models trained on all available historical data for that same cultivar. Furthermore, the authors demonstrated that simultaneously learning the related tasks of cold hardiness and budbreak provides additional accuracy benefits. In transfer learning scenarios, fine-tuning a multi-head model proved to be a highly effective strategy for new cultivars.