Project
Hybrid Machine Learning process-based modelling approaches for climate adaptation strategies
Climate change poses risks to food security by affecting various aspects of the agri-food value chain. Importantly, extreme droughts and heat waves reduce yields and nutritional value of crops, accelerate pests, diseases, and crop senescence, and disrupt supply chains and reduce shelf-life of products. Also, these impacts are higher when droughts and heat waves occur simultaneously (i.e., compound events ). Climate adaptation strategies adopted by different actors in the agri-food value chain may alleviate negative impacts of climate change. Strategies include alternative farm management, pest and disease control, the use of new genotypes or alternative crops, shifting supply lines, and post-harvest measures. The ability to forecast the effects of adaptation strategies is essential in evaluating their effectiveness.
With increasingly available data on crop phenotyping, proximal and remote sensing, transportation, delivery, consumption, etc., data-driven models based on Machine Learning (ML) are gaining importance in guiding the decision-making in various aspects of the agri-food value chain . The use of ML-based approaches for evaluating climate adaptation strategies is currently hampered by two things. Firstly, ML-based models are unreliable when extrapolating beyond the range of the training data. Extrapolation is however essential given that climate change leads to conditions that previously were rare or non-existent and of which we hence have little data. The use of hybrid approaches combining ML and process-based modelling that codify domain knowledge may remedy this drawback while using the strengths of ML and the available data . Secondly, data are scattered, noisy, and heterogeneous. The standardization of data schemata through ontologies and knowledge graphs would benefit the organization of data and provide new information for training ML-based models.
Summary
The aim of this proposal is to develop and apply hybrid approaches based on ML and process-based modelling for the assessment of climate adaptation strategies for different actors in the agri-food value chain. This project is built on the following four work packages:
WP1: data standardization through the development of ontologies and integrated knowledge graphs to harmonize terminology and data sources. The WP focused on knowledge graphs for interoperability of data from different domains.
WP2: development of hybrid ML and process-based models for crop growth from time series data that involve essential genotype-by-environment interactions to describe crop responses to environmental stresses (e.g. drought). This WP focused on hybrid ML models based on the integration of synthetic data from crop growth models (Tipstar, APSIM) and physics informed neural network to predict yield and crop stresses
WP3: pest prediction models using ML and will focus on the development of ML models for yellow stem borer infestation on rice in India.
WP4: assessment of post-harvest tomatoes quality, this WP will focus on the comparison of modelling approaches of different complexity from linear models to hybrid ML and the relation between complexity and model fitness.