Constraining Unseen Drought Impact Risks with Machine Learning and Climate Simulations |
|
Vitus Benson
,
Benjamin Stocker
,
Markus Reichstein
|
Project descriptionDrought is among the most consequential climate hazards for society, and its impacts—on crop yields, water security, ecosystem health and habitat provisioning—are rising as the climate warms[1]. A particular challenge are unseen events: droughts that are physically plausible and even likely, but that have not materialised in the recent observational record of a given place[2]. Because risk assessment, early warning and long-term planning have traditionally leaned on what a region has experienced before, they are structurally ill-equipped to anticipate impacts in places that have so far been spared. This blind spot is most acute at the decadal horizon, precisely the time scale that matters for infrastructure, land-use and adaptation planning, yet one that classical early-warning systems do not cover[3]. The core question of this PhD is therefore: can we estimate the risk of drought impacts in regions that have not seen severe droughts recently—how is that risk distributed in space, where are the hotspots, and what drives it?The project approaches drought impacts through the lens of globally available data, which is to say primarily satellite remote sensing. The premise is that the imprint of drought on ecosystems is legible in the vegetation response to weather, and that this response can be observed consistently across the globe. To estimate risk in places without a recent local analogue, the work builds on a property that modern AI models have begun to demonstrate convincingly: the ability to generalise across locations, leveraging space-for-time substitution so that an event that is unprecedented at one site can be informed by dynamically similar events seen elsewhere during training[4]. Concretely, the PhD will work with the next generation of EarthNet models[5], developed within the WeatherGenerator and EarthGenerator projects, and probe their ability to predict drought impacts under decadal scenarios—drawing on ISIMIP forcings, boosted-extremes (ensemble boosting) simulations, and climate emulators such as MESMER. By coupling these scenario ensembles with impact-aware vegetation models, the project aims to produce spatially resolved maps of decadal drought-impact risk and to identify where the emerging hotspots lie. In a second phase, the work moves from estimating risk to understanding it. The PhD will investigate the mechanisms and drivers behind these decadal drought-impact risks, with particular attention to the role of spatial context and of lateral biogeochemical and hydrological transport—processes that are often neglected when ecosystems are modelled as independent columns, yet that may decisively shape whether a landscape buffers or amplifies a drought. By interrogating what the AI models have learned and testing it against process understanding, the project seeks not only predictive skill but also explanatory insight that can inform proactive, decadal-scale adaptation and early-warning efforts. Working group & collaborationThe PhD candidate will join the newly founded NICE (Neural Inference of the Carbon Cycle and Early Warning Systems) project group at the Department of Biogeochemical Integration as part of the EarthGenerator Horizon Europe project. In addition, they will join the EarthNet team and become an external member of the Geocomputation and Earth Observation Group at Uni Bern and are expected to conduct a research visit there. The PhD candidate will be jointly supervised by Vitus Benson (MPI BGC), Markus Reichstein (MPI BGC) and Benjamin Stocker (Uni Bern).RequirementsApplications to the IMPRS-gBGC are open to well-motivated and highly-qualified students from all countries. Prerequisites for this PhD project are:
Reference
|