Estimating and Forecasting Grapevine Phenology using Deep Learning and Mechanistic models: A comparative study

Accurate monitoring of key phenological stages, such as bud-break, blooming, veraison, and berries ripe, is essential for optimizing vineyard management, especially in the context of climate change. Grapevine phenology, primarily driven by temperature, has been extensively studied with the most common prediction methods based on the growing-degree-days (GDD) approach. In this paper we propose and compare two novel methodologies, respectively based on a mechanistic and a deep learning approach, to detect and forecast key phenological stages. The mechanistic model simulates the effects of both air temperature and water availability on the major plant metabolic processes that drive the annual phenological cycle (vegetative and reproductive growth, and sugar accumulation in the berry), whereas the deep learning model, based on a Recurrent Neural Network, infers directly the phenological cycle taking as input the time series of the daily weather station measurements. Using seven years of phenological data collected from five grape varieties (‘Aglianico’, ‘Greco’ ‘Syrah’, ‘Touriga Franca’, and ‘Touriga Nacional’) across four Mediterranean sites in Italy, Spain, and Portugal, we evaluate the performance of these models against traditional GDD-based methods. Cross-validation results show that while the mechanistic model excels at predicting bud-break, with a Mean Absolute Error (MAE) of 4.9 days compared to 5.7 and 5.6 days of the GDD-based and the deep learning model, respectively, the deep learning model is an all-around optimal methodology in terms of predicting accuracy across different stages, resulting in an overall MAE slightly under 5 days, with an average error reduction of 0.6 to 1.6 days respect to the two GDD-based models tested. Furthermore, we illustrate how these models can leverage sub-seasonal weather forecasts to predict phenological events in advance, offering practical tools for vineyard managers. This was particularly true for the deep learning model, that was able to predict grapevine phenological stages two weeks in advance with an average error lower than 6 days. These findings highlight the potential of these novel approaches for automating phenology monitoring and improving decision-making in viticulture under changing climate conditions.