Transfer Learning via Auxiliary Labels with Application to Cold-Hardiness Prediction

TL;DR

The paper tackles the challenge of predicting grape cold-hardiness for cultivars lacking LTE data by introducing Transfer via Auxiliary Labels (TAL), a transfer-learning framework that leverages auxiliary phenology data. It develops two multi-task architectures (MultiHead and Embedding) and two transfer mechanisms (model selection and model averaging), with a specialized instantiation for cold-hardiness prediction. Empirical results on real-world grape datasets show that TAL, especially when using an Embedding model with model averaging and random-embedding augmentation, substantially improves LTE predictions for target cultivars relative to baselines and closes the gap toward LTE-based oracle performance. The work provides a practical, data-efficient approach for extending cold-hardiness modeling to unmodeled cultivars, with direct implications for agricultural decision support and adoption via grower-focused tools like AgWeatherNet.

Abstract

Cold temperatures can cause significant frost damage to fruit crops depending on their resilience, or cold hardiness, which changes throughout the dormancy season. This has led to the development of predictive cold-hardiness models, which help farmers decide when to deploy expensive frost-mitigation measures. Unfortunately, cold-hardiness data for model training is only available for some fruit cultivars due to the need for specialized equipment and expertise. Rather, farmers often do have years of phenological data (e.g. date of budbreak) that they regularly collect for their crops. In this work, we introduce a new transfer-learning framework, Transfer via Auxiliary Labels (TAL), that allows farmers to leverage the phenological data to produce more accurate cold-hardiness predictions, even when no cold-hardiness data is available for their specific crop. The framework assumes a set of source tasks (cultivars) where each has associated primary labels (cold hardiness) and auxiliary labels (phenology). However, the target task (new cultivar) is assumed to only have the auxiliary labels. The goal of TAL is to predict primary labels for the target task via transfer from the source tasks. Surprisingly, despite the vast literature on transfer learning, to our knowledge, the TAL formulation has not been previously addressed. Thus, we propose several new TAL approaches based on model selection and averaging that can leverage recent deep multi-task models for cold-hardiness prediction. Our results on real-world cold-hardiness and phenological data for multiple grape cultivars demonstrate that TAL can leverage the phenological data to improve cold-hardiness predictions in the absence of cold-hardiness data.

Figures (4)

• Figure 1: Cold-hardiness data for three grape cultivars compared to minimum air temperature. The data points are $LTE_{50}$, or the temperature at which 50% of the grape buds will experience lethal freezing (see Section \ref{['sec:related']}). Note how the cold-hardiness has a general seasonal trend as well as responses to the local weather. Different cultivars show distinct cold-hardiness responses, indicating that growers need models that are tuned for their individual cultivars.
• Figure 2: Overview of TAL. We learn a model for the source cultivars using the corresponding primary LTE and auxiliary phenology data. The new target cultivar only has phenology data, so we cannot directly learn a target LTE model and instead transfer knowledge from the source cultivar models.
• Figure 3: (left) Diagram of the backbone encoder model $f_{\theta_0}$ used for cold-hardiness prediction. Since the inputs for cold-hardiness prediction are time series, this encoder is a recurrent network that encodes any time series into a fixed length vector. (middle) The MultiHead MTL architecture can use any form of backbone encoder to produce primary and auxiliary outputs for each source task. (right) The Embedding MTL architecture concatenates the input with a learned task specific embedding vector and passes this concatenation to encoder and feeds the output to the shared prediction head.
• Figure 4: Comparison of different embedding sets (S, CR, LR-3 and LR-17). The green lines show the LTE predictions for the 2020 season using embeddings from a model trained with Barbera as the target task and all others as the source tasks. The black dots indicate the Barbera ground truth LTE values.