Advisor: Wouter Arnoud Dorigo
673 million people faced hunger in 2024, while climate change and a growing world population put further pressure on global agriculture. Given that 13% of global food production goes to waste in the supply chain, optimised resource management is essential to reduce both food waste and hunger. Crop yield forecasts enable decision-makers to better manage stocks and guide the import and export of crops. Despite enormous efforts over the past decades to provide reliable crop yield forecasts, their accuracy is often inadequate due to data scarcity and the underestimation of the impact of extreme events, such as droughts, on crop yields. Currently used models - both machine learning and process-based - perform insufficiently when too little crop yield data or predictor data is available for a certain region or conditions. Therefore, in this thesis, I focused on three research question:

• How do machine learning and EO based crop yield forecasting systems perform in drought years?
• How can we improve crop yield forecasting in data-scarce situations?
• Can we use this information to also estimate yield losses caused by other extremes, like war-related yield losses?

In the first paper of this thesis, I demonstrated how drought-related crop yield losses can be detected early but remain underestimated in current crop yield forecasting systems. Thus, I concluded that additional data from drought years is required to accurately capture the impacts of droughts on crop yields. In the second paper, I showed how transfer learning — a novel machine learning technique that trains and tests a model across different domains — can address the issue of data scarcity. Specifically, I trained a field-scale crop yield forecasting model at the regional scale and fine-tuned it using only a limited number of field-scale samples. In the third paper, I returned to forecasting crop yields during extreme events: estimating war-related crop yield losses in Ukraine using transfer learning. Although crop yield losses remained underestimated, transfer learning led to significantly improved crop yield forecasts in both average years and during the war.
Going forward, I identified two main steps to further advance crop yield forecasting: (1) expand the use of transfer learning for crop yield forecasting in any data-scarce situation, and (2) further improve forecasting during extreme events such as droughts and conflicts. Through this thesis, I contributed to improved crop yield forecasting in data-scarce situations and during extreme events, which will allow decision-makers to better manage stocks and guide imports and exports to reduce hunger in a changing climate.