Calibrating Biophysical Models for Grape Phenology Prediction via Multi-Task Learning
William Solow, Sandhya Saisubramanian · 2025
Read Calibrating Biophysical Models for Grape Phenology Prediction via Multi-Task Learning by William Solow, Sandhya Saisubramanian (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2508-03898.txt.
Introduction
William Solow and Sandhya Saisubramanian improve grape growth-stage forecasts in a 2025 paper. Knowing when vines bud, flower, and ripen guides watering and feeding choices. Classic physics-based models use past field data but lack fine precision. Deep learning could help, yet data per grape variety is sparse.
Their hybrid joins multi-task learning with a recurrent network. The learner predicts the settings of a physics model that can be tuned by gradients. Sharing across varieties adds strength while the physics keeps results sound. Tests on real and synthetic data beat both classic models and plain deep learners. Other crop traits, such as cold-hardiness, are also forecast well.
Catalog record
| CREATOR | William Solow, Sandhya Saisubramanian |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 47 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2508-03898
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Calibrating Biophysical Models for Grape Phenology Prediction via Multi-Task Learning William Solow, Sandhya Saisubramanian Abstract Accurate prediction of grape phenology is essential for timely vineyard management decisions, such as scheduling irrigation and fertilization, to maximize crop yield and quality. While traditional biophysical models calibrated on historical field data can be used for season-long predictions, they lack the precision required for fine-grained vineyard management. Deep learning methods are a compelling alternative but their performance is hindered by sparse phenology datasets, particularly at the cultivar level. We propose a hybrid modeling approach that combines multi-task learning with a recurrent neural network to parameterize a differentiable biophysical model. By using multi-task learning to predict the parameters of the biophysical model, our approach enables shared learning across cultivars while preserving biological structure, thereby improving the robustness and accuracy of predictions. Empirical evaluation using real-world and synthetic datasets demonstrates that our method significantly outperforms both conventional biophysical models and baseline deep learning approaches in predicting phenological stages, as well as other crop state variables such as cold-hardiness and wheat yield. Code — https://tinyurl.com/IRAS-DMC-MTL Figure 1: Overview of the problem and solution approaches. Left: Historical weather and phenology observations are available on a per-cultivar basis to calibrate phenology models. The biophysical, deep learning, and hybrid models predict the onset of the phenological stage. Red ‘x’ marks denote predictions errors greater than five days and green check marks denote errors less than five days. A question mark represents a biologically inconsistent prediction. Biophysical models make significant prediction errors and deep learning methods produce biologically inconsistent predictions. Our proposed hybrid approach improves prediction accuracy substantially. Right: Illustrations of current and proposed solution approaches for crop state predictions. Introduction Seasonal vineyard tasks such as fertilization, irrigation, pruning, and harvesting rely on accurate predictions of grape phenology ( Keller et al. 2016 ; Keller and Hrazdina 1998 ; Milani and Cawley 2024 ; Zapata et al. 2017 ) . The key grape phenological states are: bud break, bloom, and veraison. Inaccurate predictions of these states can lead to poorly timed interventions, resulting in reduced yield, quality, and lowered vineyard health ( Balint and Reynolds 2017 ) . However, accurate phenology prediction is challenging due to (1) limited availability of historical data for per-cultivar calibration and sparse observations during each growing season ( Zapata et al. 2017 ) , and (2) the need to accurately model complex relationships between daily weather features and phenology ( Badeck et al. 2004 ) . Existing approaches to this problem typically fall into two categories: mechanistic biophysical models and data-driven deep learning approaches. The Growing Degree Day (GDD) model is a widely used biophysical model for phenology prediction, that predicts phenological stages based on daily accumulated heat units ( Ortega-Farias and Riveros-Burgos 2019 ) . The GDD model is calibrated on a per-cultivar basis using multivariate regression over historical field observations. Despite the availability of many weather features, the GDD model only uses ambient temperature as input, limiting its expressiveness ( Badeck et al. 2004 ) . Despite various proposed improvements ( Zapata et al. 2017 ; Fraga et al. 2016 ; Garcia de Cortazar Atauri et al. 2017 ) , the accuracy of GDD-based phenology predictions remains relatively low and vineyards are actively seeking improvements ( Reynolds 2022 ) . Deep learning methods offer a promising alternative due to their ability to model complex, nonlinear relationships between weather variables and phenological stages. A recent work proposed a multi-task classification model to predict grape bud break , a key phenological stage ( Saxena et al. 2023a ) . They leveraged shared information across cultivars to improve accuracy, particularly for data-scarce cultivars. However, this approach often produced biologically inconsistent predictions, such as predicting bud break followed by a return to dormancy within a few days, which violates the unidirectional progression of phenological stages. Such inconsistencies introduce ambiguity into prediction interpretation and make them unsuitable for decision-making in the field, especially since grape growers rely on phenology models to make medium range (7–14 day) operational forecasts ( Reynolds 2022 ) . An emerging direction is hybrid modeling , which combines deep learning and biophysical models to address their respective limitations and improve prediction accuracy. It has recently been applied to accurately predict bloom date in cherry trees by leveraging deep learning to approximate the internal temperature response function within the GDD model ( Van Bree, Marcos, and Athanasiadis 2025 ) . However, this formulation does not consider the effects of exogenous weather features, such as solar irradiation and rainfall, which significantly affect phenological development and are particularly important for season-long grape phenology forecasting ( Badeck et al. 2004 ) . To address the challenges in accurate grape phenology prediction, we propose a hybrid modeling approach that uses a recurrent deep learning model to predict daily parameters of the GDD model conditioned on exogenous weather features. Our approach uses multi-task learning via a per-cultivar embedding to efficiently share data among cultivars and increase per-cultivar prediction accuracy. We provide a novel implementation of the GDD model in a differentiable framework ( Paszke et al. 2017 ) to enable gradient descent on the neural network parameters with supervised learning. With the daily dynamic parameterization of the GDD model, our proposed hybrid approach produces accurate phenological forecasts and can be extended to other domains such as grape cold-hardiness and wheat yield ( Ferguson et al. 2011 ; Ceglar et al. 2019 ) , as shown in our experiments. Figure 1 illustrates the problem and our proposed approach. Our main contributions include: (1) presenting a novel hybrid approach for accurate phenological forecasts by predicting parameters of the biophysical model conditioned on the weather features; (2) formulating the prediction problem as a multi-task learning problem that leverages data efficiently across grape cultivars; and (3) empirical evaluations using real-world and synthetic datasets that demonstrate our approach’s robustness and increased accuracy compared to state-of-the-art biophysical baselines, deep learning approaches, and hybrid models. We evaluate our approach on three prediction tasks: predicting grape phenology using real-world data, cold-hardiness using real-world data, and seasonal wheat yield using synthetic data. The two additional domains, cold-hardiness and wheat yield prediction, share the difficulties raised by grape phenology prediction of sparse per-cultivar data and strict biological structure. Figure 2: Daily crop state observations for five cultivars of (a) grape phenology (b) grape cold-hardiness and (c) wheat yield during a single growing season. Modeling approaches must predict these curves with biologically consistency. Despite experiencing the same weather, cultivars exhibit different behaviors, making naive data aggregation inadequate and motivating the use of a multi-task approach. Background and Related Work Model Calibration in the Agricultural Sciences Before a biophysical model is used for agricultural crop state prediction, it must be calibrated with historical data. Common approaches used in the agricultural community for parameter calibration include brute force search ( Ferguson et al. 2014 ) , regression techniques ( Zapata et al. 2017 ) , and Bayesian optimization ( Seidel et al. 2018 ) . However, these approaches assume that a stationary parameter set best explains the observed time series data during the growing season. Given the simplicity of the GDD grape phenology model, this assumption may not hold in practice. In contrast, our