Model generalization hackathon: Projecting future climate impacts to crop yields

[lilybellesweet]

Title

Responsible person(s)

Lily-belle Sweet, Department of Compound Environmental Risks, Helmholtz Centre for Environmental Research (UFZ), (lily-belle.sweet@ufz.de)
Daniel Klotz, Department of Compound Environmental Risks, Helmholtz Centre for Environmental Research (UFZ), (daniel.klotz@ufz.de)
Brian Groenke, Alfred Wegener Institute

Format

Hackathon

Timeframe

Full day

Description

Can we learn from the recent past to predict climate impacts in the future?

Machine learning models are frequently trained on observed data from the last decades and then used to make projections of future climate change impacts. However, the ability of such models to generalise to these unseen conditions outside of the observed distribution is not guaranteed. How far into the future can we make good predictions? Which types of models or training methods do better or worse? Can domain generalisation strategies help, and if so, how much? We have created a benchmark dataset to help answer these questions, using simulated agricultural maize and wheat yields from biophysical crop models.

Participants will train models to predict end-of-season annual gridded yields for current global cropping areas using data from 1980-2020, using daily growing-season climate data (precipitation, solar radiation, temperature), soil texture and CO2 concentration at 0.5 degree spatial resolution. The models will then be evaluated based on their ability to predict yields from 2020 to 2100 under a high-emissions climate change scenario (RCP8.5).

No climate/agricultural knowledge is needed to participate and the data will be processed to be very straightforward to work with. It's a large, multivariate and high-dimensional dataset, so there is a lot of scope to experiment with interesting model architectures.

Timeline: The day will start with an introduction to the problem, including a quick explainer of the life cycle of maize and wheat and how their growth is affected by weather, followed by a short hands-on tutorial for downloading the data (probably via Kaggle) and using it to train a simple model. After this, participants are free to work individually or in teams to train their models and submit their predictions. At the end of the day, we invite each team/participant to briefly present what they did and share their experiences. We will show not just an overall leaderboard, but also a breakdown of scores for each decade.

Further outcomes: This event would be part of a broader activity organised by the AgMIP Machine Learning team AgML, and the challenge will remain online for participation for ~six months outside of this satellite event. We aim to then publish the dataset as a benchmark for model generalisation, along with some analysis of the results (by breaking down the test scores based on different types of model architectures, modelling methodology choices etc). Hackathon participants may be able to contribute to this if interested, but would also be free (and encouraged) to publish their results independently if they are able to develop particularly high-performing models.

Requirements

Room with presentation equipment, enough power sockets for all participants, good WiFi connection for downloading the data (although participants can also download in advance)
10+ attendees
Access to the Jülich cluster for participants would be nice, but not a requirement

[Divya1205]

Hi Lily, Is it possible to participate virtually?