PhD defence
Machine learning for large-scale crop yield forecasting
Summary
Accurate crop yield forecasts are valuable to adjust food security policies, farm management practices and marketing plans for farm products. Operational large-scale yield forecasting systems do not currently use machine learning. Machine learning can combine domain knowledge with data-driven learning and provide increased automation for large scale. This thesis investigated the benefits and challenges of using machine learning for large-scale crop yield forecasting. It serves as a roadmap by identifying key requirements (the “what”) and designing workflows to address them (the “how”). The workflows had to be automated to produce consistent and reproducible yield forecasts across multiple spatial levels (e.g. grids, regions). They also had to account for data sparsity and produce forecasts backed by explanations understandable to human stakeholders. The four sub-objectives addressed these requirements and, in the process, made scientific contributions related to defining the forecasting setup, handling missing data and evaluating interpretability of model forecasts.