tl; dr:
● phenology describes the periodic life cycle of a plant
● phenology varies between different varieties, i.e.
Merlot takes longer to ripen than Pinot noir
● phenology is mostly driven by temperature
● phenology models are currently built for specific varieties (or for an average variety)
● problem: Variety-specific are needed but tedious to implement into real world applications.
● challenge: Develop a model that is able to predict the phenology for multiple varieties using a
dynamical systems approach and machine learning.
Challenge 1: Develop a grapevine phenological
model that works for all grapevine varieties
Phenology
Bud break (spring) Bloom (summer) Maturation (autumn)
Current models
Degree day models
Accumulated temperature (often with thresholding) is correlated with BBCH stages.
Phenology encoded as BBCH stages
Examples:
BBCH 09: bud break
BBCH 11-19: leaf development BBCH 60-69: bloom
BBCH 80-89: maturation
Source: Molitor et al (2020)
Challenge
The one and only phenology model
Develop a model that is able to predict the phenology for multiple varieties using a dynamical systems approach and machine
learning.
Challenge
Phenology as dynamical system
State, i.e. BBCH
Dynamics Time
Control, i.e. temperature and variety
Parameters, i.e. weights and biases of a neural network
Challenge
Learning dynamics of the Lorenz System (Raissi et al. 2018)
Why is this important?
Downy mildew Powdery mildew
Why is this important?
Resources for the challenge
Literature
Phenology:
Leolini, Luisa, et al. "Phenological model intercomparison for estimating grapevine budbreak date (Vitis vinifera L.) in Europe." Applied Sciences 10.11 (2020): 3800.
Molitor, Daniel, Helder Fraga, and Jürgen Junk. "UniPhen–a unified high resolution model approach to simulate the phenological development of a broad range of grape cultivars as well as a potential new bioclimatic indicator."
Agricultural and Forest Meteorology 291 (2020): 108024.
Modelling nonlinear dynamics with neural networks:
Qin, Tong, Kailiang Wu, and Dongbin Xiu. "Data driven governing equations approximation using deep neural networks." Journal of Computational Physics 395 (2019): 620-635.
Raissi, Maziar, Paris Perdikaris, and George Em Karniadakis.
"Multistep neural networks for data-driven discovery of nonlinear dynamical systems." arXiv preprint arXiv:1801.01236 (2018).
Data
Phenology:
https://data.pheno.fr/
Weather:
https://cds.climate.copernicus.eu/cdsapp#!/dataset/reanalysis-er a5-land