PaperlayerView exact revision on arXiv

Source-linked AI summary

Cross-Regional Grapevine Cold Hardiness Prediction via Learned Multimodal Latent Representations

William Solow, Paola Pesantez-Cabrera, Markus Keller, Lav Khot, Sandhya Saisubramanian, Alan Fern

arXiv:2608.31097v1cs.AI

TL;DR

Cold-hardiness models are accurate locally but lack principled transfer to data-scarce regions and new cultivars. The paper learns transferable latent representations and infers them from text or limited observations, achieving stronger cross-region performance while exposing biological-realism limitations in unconstrained deep learning.

  • Problem

    Local cold-hardiness models and scarce LT50 data limit reliable prediction transfer to new regions and cultivars.

  • Method

    A multi-task framework learns region- and cultivar-related latent representations, approximated for unseen tasks using text descriptions or limited historical observations.

  • Results

    Our DL and Hybrid models outperform the Ferguson model by 1°C RMSE in data-limited regions, while text transfer outperforms all baselines.

  • Takeaways & Limitations

    Cross-region data sharing and transferable latent representations improve cold-hardiness prediction, including in data-limited regions.

  • Takeaways & Limitations

    The DL model produces biologically unrealistic predictions, including failure to reach known spring minima or decreases after warming.

Abstract

from arXiv · show

Accurate daily predictions of cold hardiness in woody plants are critical in regions where freezing temperatures can damage dormant buds and reduce seasonal yield. Existing biophysical, hybrid, and deep learning models have shown high predictive accuracy when trained on local data but remain largely site-specific. The limited availability of cold hardiness data, coupled with the lack of principled methods for transferring cold hardiness predictions to new regions and cultivars, has limited the broader adoption and practical utility of these approaches, particularly in data-scarce regions. To address these limitations, we propose a cold hardiness prediction framework that learns a transferable latent representation by capturing region-specific variation through learned embeddings. To enable prediction in previously unseen regions, we infer embeddings from (1) text descriptions of the cultivar and growing region, and (2) limited historical observations, supporting both zero-shot and few-shot transfer. Experiments on datasets from six regions across North America demonstrate that our approach consistently outperforms state-of-the-art cold hardiness prediction methods, yielding more accurate predictions and substantially improving transfer to data-scarce regions.

Introduction

Cold hardiness prediction matters for costly frost-management decisions, but direct LT50 monitoring is impractical and existing models generalize poorly beyond their training sites or cultivars. The framework addresses this gap by learning transferable regional representations and inferring them from text or limited historical observations.

  • Motivation: LT50 is the lethal temperature at which 50% of dormant grapevine buds freeze, making accurate predictions important for precision frost management.Preventive measures such as wind fans or kerosene heaters are costly and should be used only when necessary.
  • Motivation: Frequent LT50 measurement is impractical because it requires specialized equipment, trained technicians, and bud removal before damage can be assessed.Consequently, growers cannot monitor vine cold hardiness in real time.
  • Research gap: Limited per-cultivar data makes diverse multi-region datasets difficult to curate, so existing models often rely on site- or cultivar-specific training data.This reliance leads to poor out-of-region generalization.
  • Research gap: The same cultivar can respond differently across regions because of unobserved factors including soil properties, local climate, and management practices.These regional differences make naive prediction transfer challenging.
  • Contribution: The proposed framework learns region-level latent representations across regions, stations, vineyards, and cultivars, then infers them from text or limited observations for zero-shot and few-shot transfer.It supports unified comparison across biophysical, deep learning, and hybrid models.

Related Work

Prior cold-hardiness models provide complementary strengths but remain limited by site- or cultivar-specific assumptions and do not support prediction transfer to new regions. The paper positions transfer learning as a way to account for regional variance while comparing modeling paradigms systematically.

  • Cold Hardiness Modeling: Grapevine cold hardiness is dormant-bud tolerance to freezing without lethal tissue damage and changes during dormancy in response to temperature and other latent variables.Its measurement has required extensive data collection using differential thermal analysis.
  • Existing Models: Ferguson and GrapeHardiNet make site- and cultivar-specific predictions, while NYUS2 aggregates regions but does not account for regional variance.These approaches therefore differ in how they share information across locations and cultivars.
  • Transfer Gap: Biophysical, deep learning, and hybrid approaches do not support transfer when local training data are required for LT50 prediction.Aggregating across regions also violates assumptions that hold for site-specific modeling.
  • Transfer Gap: The paper presents transfer learning for LT50 prediction as a response to regional variance and identifies it as the first such study to its knowledge.This extends multi-task learning beyond site-specific aggregation.

Problem Formulation

The paper formulates cold-hardiness prediction as multi-task sequence prediction, where weather and task identity jointly determine daily LT50 values. A shared latent task representation unifies biophysical, deep learning, and hybrid models and supports transfer across regions and cultivars.

  • Problem Formulation: Each prediction task is a unique region, weather station, vineyard, and cultivar combination sharing biological processes but differing in environmental and cultivar-specific conditions.Learning across related tasks can improve accuracy and facilitate transfer to new regions and cultivars.
  • Framework: The framework combines a task encoder ϕ, a weather-and-state function ψ, and a biophysical model M.Together, these components represent task information, temporal relationships, and LT50 generation.
  • Framework: The composed prediction is ŷ_i = M[ψ[ϕ(r, s, v, c), w_i], w_i], and end-to-end training learns task representations used for transfer.The learned latent vector ϕ is conditioned on region, station, vineyard, and cultivar.
  • Unifying Modeling Lens: Biophysical, deep learning, and hybrid models are expressed as instances of one multi-task formalism, enabling systematic comparison and transferable latent representations.Their differences concern task-specific representation and the mapping from weather to LT50.

Transfer Learning of LT50 Predictions

The transfer approach approximates latent task representations for unseen regions and cultivars using text descriptions or limited historical observations. Learned representations capture regional variation, and text-based transfer provides a zero-shot route while observation-based transfer supports few-shot adaptation.

  • Transfer Motivation: Averaging predictions across training regions ignores regional variation and discards target-region observations or priors.The proposed approach instead conditions predictions on a learned latent vector for each task.
  • Latent Features: Varying the learned representation while holding weather fixed produces substantially different LT50 predictions for 12 cultivars.This indicates that the representation captures cold-hardiness factors beyond the observed weather.
  • Latent Features: t-SNE embeddings cluster primarily by region rather than cultivar, suggesting regional characteristics strongly influence cold hardiness.Vineyards in the same region receive similar color coding in the visualization.
  • Transfer Strategy: The two transfer methods approximate ϕ for unseen regions and cultivars without retraining the LT50 model for each new task.This enables a wider range of cold-hardiness responses at new vineyards.
  • Transfer via Text Embeddings: Text descriptions of cultivars, vineyards, weather stations, and regions are encoded into embeddings to approximate ϕ when prior target-region data are unavailable.Descriptions can be prepared with domain experts or scientific literature and encoded using E5.
  • Transfer Using Limited Data: Limited historical LT50 and weather observations can also approximate ϕ for few-shot transfer, despite irregular sampling within daily weather sequences.The method uses cross attention to align LT50 and weather sequences before producing a latent vector.

Experiment Setup

The experiments compare unified deep learning and hybrid models with deployed baselines across six North American regions, then test transfer to unseen regions and cultivars using text and historical observations.

  • Datasets: Six real-world datasets span BCOV, ONNP, Washington, New York, Michigan, and Nova Scotia across 12 cultivars.Weather data include daily minimum, maximum, and average temperatures and precipitation from nearby open-field stations.
  • Models: The proposed framework is evaluated as Deep Learning and Hybrid models against Ferguson, GrapeHardiNet, and NYUS 2.2 baselines.The framework uses Ferguson as the biophysical component M and learns latent features for regions, stations, vineyards, and cultivars.
  • Evaluation protocol: Testing uses two seasons per cultivar and vineyard, validation uses one season, and remaining observations are used for training.Hyperparameters and model selection are determined using the validation set.
  • Transfer experiments: Regional transfer trains on BCOV and ONNP, then evaluates on four other regions; cultivar transfer withholds one cultivar across all regions.Transfer approximators infer latent representations from text descriptions or limited historical seasons after the multi-task model is frozen.
  • Evaluation protocol: Each model is trained five times with different seeds, and performance is reported as average daily RMSE in degrees Celsius.RMSE is computed across seasons and cultivars by region or across each cultivar.

Results and Discussion

Across regions and transfer settings, the proposed models improve predictive accuracy over deployed baselines, while the Hybrid model produces substantially fewer biologically unrealistic predictions than the DL model.

  • Overall performance: 9% average improvement over the next best deployed model is achieved by both general models across six regions without transfer.Paired t-tests across evaluation seasons confirm that most improvements over the best-in-region deployed model are statistically significant at p < 0.05.
  • Overall performance: 1°C RMSE improvement over Ferguson is achieved by the DL and Hybrid models in data-limited New York, Nova Scotia, and Michigan, equivalent to a near 30% increase in performance.Aggregating data also improves performance over the site-specific GrapeHardiNet model in BCOV, ONNP, and Washington.
  • Regional transfer: Text-based transfer outperforms all baselines on four unseen regions and matches the all-region DL model in some regions.Few-shot transfer performs slightly worse, while naive cultivar averaging remains a strong prior and direct regional embeddings perform poorly.
  • Cultivar transfer: Text descriptions are more accurate than naive averaging for withheld-cultivar transfer, while Ferguson performs poorly under transfer.The base embedding approach is comparable for cultivar transfer, suggesting uninformative cultivar representations are less damaging than uninformative regional representations.
  • Ablation and realism: More verbose text embeddings and additional years of data are evaluated as auxiliary information for transfer accuracy and biological realism.Biological unreality is defined using spring LT50 behavior, including failure to reach the cultivar’s known minimum or decline during warming temperatures.
  • Biological realism: 6.2% versus 1.9% biologically unrealistic predictions occur for the DL and Hybrid models, respectively, across all cultivars.The DL model’s unrealistic-prediction rate increases more under transfer, motivating deployment of the Hybrid model despite slightly lower accuracy.

Summary and Future Work

The framework addresses the inability of current platforms to predict for unseen cultivars and regions by offering transfer methods with limited or no prior data. It is deployed in collaboration with project partners.

  • Practical scope: Current models on AgWeatherNet and NEWA cannot predict for cultivars and regions outside their available training data.The proposed methodology provides multiple transfer routes for unseen regions and cultivars with limited or no prior data.
  • Practical scope: The Hybrid model is currently deployed on AgWeatherNet.The deployment reflects the framework’s practical relevance for operational cold-hardiness prediction.
  • Study scope: The study uses six diverse cool-climate viticultural regions selected for available data, vineyards, cultivars, and seasons.The regions are British Columbia, Ontario, Washington, New York, Michigan, and Nova Scotia.
  • Study scope: Daily weather observations from nearby open-field stations provide temperature and precipitation inputs alongside LT50 observations.Station sources include AgWeatherNet, Canadian government stations, and NOAA stations.

Appendix A.2 Data Progressing Procedure

The procedure filters and repairs weather data, retains seasons with valid LT50 observations, and acknowledges measurement variability across regions and laboratories.

  • Seasons with more than 10% missing temperature features are discarded; otherwise, missing values are linearly interpolated and weather features are z-score normalized.
  • Seasons with at least one valid LT50 observation are included, while missing LT50 days are masked during loss computation and backpropagation.
  • LT50 samples are noisy because multiple dormant buds are frozen simultaneously during differential thermal analysis.
  • Regional laboratories use slightly different differential thermal analysis procedures, which may contribute to larger standard deviations in the results.

Appendix B.1: Evaluation Protocol

The evaluation withholds seasons by cultivar, vineyard, and region across five seeds, while transfer experiments use British Columbia’s Okanagan Valley and Ontario’s Niagara Peninsula as training regions.

  • Across five seeds, two seasons per cultivar, vineyard, and region are withheld for testing, with an additional season per cultivar withheld from training.
  • The most data-scarce setting leaves one training season and one validation season, while other cultivars receive at least two training seasons.
  • BCOV and ONNP serve as transfer-training regions because their weather stations cover multiple vineyards and provide more diverse representations.
  • Figure 7 depicts the backbone, latent-representation interactions, text embeddings, and few-shot cross-attention components of the transfer framework.

Appendix B.2: Model Architecture

The architecture combines recurrent sequence modeling with learned representations for regional and cultivar variation, supporting direct LT50 prediction, hybrid biophysical prediction, and transfer through text or historical data.

  • Model Architecture: A GRU-based backbone models the complex temporal relationship between weather and cold hardiness, with linear layers before and after the recurrent network.
  • Model Architecture: Cultivar, vineyard, weather-station, and region representations are concatenated with daily weather features before entering the model backbone.
  • Model Architecture: The function approximator outputs either daily LT50 values or daily biophysical-model parameters, using tanh scaling to [−1, 1] in the hybrid case.
  • Evaluation: The model evaluations use five-fold validation and grid search over GRU size, learning rate, batch size, and learning-rate annealing.
  • Transfer Mechanisms: Text transfer constructs cultivar and regional descriptions from expert-identified attributes and encodes them with sentence embeddings.

Transfer via Few Seasons of Data

Few-season transfer estimates latent representations from limited historical LT50 observations, but irregular and variable sampling makes those embeddings unreliable; text transfer performed better in the reported results.

  • Historical LT50 observations can approximate the latent representation for a new region or cultivar, enabling few-shot transfer.
  • Washington Merlot observations spanning five dormant seasons were collected aperiodically and showed high variance.
  • Regional differences in sampling rates make transfer more challenging when a new sampling period is encountered.
  • Text transfer outperformed few-season transfer, suggesting that text embeddings provided a strong prior over the latent representation in the multi-task model.

Appendix E: Additional Experiments

Additional experiments examine biological realism, latent-representation behavior, data aggregation, and transfer performance for the framework's DL and Hybrid models.

  • Biological Realism: The DL model can produce biologically unrealistic early- and late-season LT50 predictions, whereas the Hybrid model remains biologically realistic throughout the season.The Hybrid model was deployed despite the DL model's slightly higher testing accuracy.
  • Data Aggregation Under Varying ϕ: The best DL performance uses all latent features and aggregated data, while adding data without complementary regional features generally reduces performance.The comparison evaluates aggregation across cultivars, vineyards, weather stations, and regions.
  • Interpolation of ϕ: 20% to 40% increases in RMSE occur for most cultivars as embeddings are linearly interpolated toward Cabernet Sauvignon, showing that individual latent representations affect prediction accuracy.The interpolation evaluates predictions across all six regions and indicates that accurately approximating ϕ is important for transfer.
  • Hybrid Model Transfer: The text-description transfer method outperforms other approaches and baselines in the Hybrid model, although its advantage is smaller than in the DL model.The Hybrid model performs slightly worse than the DL model, while its Base Embedding baseline performs comparatively better.
  • Hybrid Model Transfer: The transfer evaluation measures RMSE on four regions unseen during training and compares two transfer methods with multiple baselines across five seeds.The baselines include a Hybrid model trained on all regional data, Ferguson, naive averaging, and direct embeddings.
Loading 2608.31097v1…