by Lily-belle Sweet, Helmholtz Centre for Environmental Research (UFZ)

Existing crop models underestimate the impacts of individual climate extremes on yields, but agricultural yield shock can also be caused by the compounding effect of multiple, more moderate, weather events. My PhD research, supervised by Jakob Zscheischler at the Helmholtz Centre for Environmental Research (UFZ), is centred on the use of interpretable or explainable Machine Learning (ML) methods to identify interacting climate drivers of agricultural yield failure, both in process-based crop models and observation.

As well as using interpretable ML to analyse and intercompare existing agricultural models, ML can also be used to generate or gap-fill input data required for driving the models, to post-process simulations (for example, by downscaling) or to replace components of agricultural models in hybrid modelling approaches. These approaches could improve our ability to capture the complex effects of compound events on agricultural systems, and better understand how they will be affected under climate change. However, advances in this intersection of disciplines require both ML expertise and deep domain knowledge.

For my STSM within the COST action DAMOCLES, I was able to visit Alex Ruane and the Climate Impacts group at the NASA Goddard Institute for Space Studies in New-York City from January to February 2023. The aim of the visit was to review the current state-of-the-art and identify opportunities in research combining ML methods with current process-based crop modelling approaches, and, in collaboration with the Agricultural Modelling Intercomparison and Improvement Project (AgMIP), to establish a community of practice for ML in agricultural modelling. As part of the visit, we organised a one-day workshop at NASA GISS along with Ioannis Athanasiadis (WUR) to discuss these ideas and begin outlining our goals for the team.

My last day visiting the Climate Impacts group at NASA 

As I flew back to Germany, I was excited, but also nervous; would other researchers find our ideas compelling, and be interested in joining the collaborations we envisioned? The STSM was an incredible opportunity to learn from so many great researchers from multiple disciplines – this type of interdisciplinary connection and knowledge-sharing is exactly what I hoped to facilitate. As it turned out, thanks to the support and mentorship of Alex, Ioannis and Jakob, the community has grown and flourished more than I could have imagined.

The deep discussions and takeaways from the STSM were used to inform our planning of several ML-focused sessions at the 9th AgMIP Global Workshop which took place at Columbia University in June. Since then, the community (AgML) has been meeting every two weeks via Zoom and working together on coordinated activities to create carefully-curated benchmark datasets and robustly evaluate the utility of different ML methods for various agricultural modelling applications. In late January 2024, we organised the first AgML workshop at Wageningen University, with around 35 in-person attendees from all around the world. We are actively working on two collaborative activities, with several more in the early stages of development.

The first AgML workshop in Wageningen, January 2024

In the study of compound events, researchers often work with large model simulations, sometimes from multiple impact sectors. Having a closer connection with the crop modellers and scientists from other disciplines involved in the production of these datasets has widened my scientific perspective and facilitated my research enormously. I hope to maintain, and strengthen, these links between Machine Learning, agricultural modelling and compound events communities in the future.

Edited by Pauline Rivoire. Photo credits: Lily-Belle Sweet (top picture); AgML and Ioannis Athanasiadis (bottom picture).

Leave a Reply

This site uses Akismet to reduce spam. Learn how your comment data is processed.