Project

Back to overview

Machine learning for detecting compound climate drivers of extreme impacts

English title Machine learning for detecting compound climate drivers of extreme impacts
Applicant Zscheischler Jakob
Number 189908
Funding scheme COST (European Cooperation in Science and Technology)
Research institution Klima- und Umweltphysik Physikalisches Institut Universität Bern
Institution of higher education University of Berne - BE
Main discipline Other disciplines of Environmental Sciences
Start/End 01.05.2020 - 31.10.2022
Approved amount 182'492.00
Show all

All Disciplines (2)

Discipline
Other disciplines of Environmental Sciences
Information Technology

Keywords (4)

compound events; deep learning; machine learning; vegetation modelling

Lay Summary (German)

Lead
Eine sehr ungünstige Kombination von Klimaereignissen kann zu extremen negative Auswirkungen auf die Umwelt führen. Zum Beispiel kann ein warmer Winter die Bildung von Schädlingen erleichtern, die sich dann in einem feuchten Frühling stark ausbreiten und zu extremen Ernteausfällen im Sommer führen. Für viele solche extremen wetterverursachten Auswirkungen ist unbekannt, welche Kombination von Ereignissen sie verursacht haben.
Lay summary

Inhalt und Ziel des Forschungsprojekts

In diesem Projekt werden wir Methoden aus dem Maschinellen Lernen anwenden und erweitern, um die Kombination von Wetterereignissen zu delektieren, die zu extremen Auswirkungen führt. Für die Entwicklung der Methoden verwenden wir sehr lange Modellläufe eines Vegetationsmodells (LPX), welches mit modelliertem Klima aus einem globalen Klimamodell angetrieben wird. Die extremen Auswirkungen sind daher in diesem Fall extrem niedriger Kohlenstoffaufnahme der Pflanzen. Wir werden die entwickelte Methodik auch für andere Ereignisse anwenden, zum Beispiel um die Niederschlagsmuster zu identifizieren, die zu den grössten Fluten führen.


 

Wissenschaftlicher und gesellschaftlicher Kontext des Forschungsprojekts

Das Identifizieren von mehreren Wetterereignissen, die in Kombination eine extreme Auswirkung erzeugen ist für eine Reihe von Wetterauswirkungen relevant, so zum Beispiel um Fluten, Ernteausfall und Feuer besser vorherzusagen und zu modellieren. Unsere methodische Entwicklung wird auch für andere Forschungsfelder und Institutionen anwendbare sein, deren Ziel es ist, bessere Abschätzungen für wetterbezogene Risiken zu machen.

Direct link to Lay Summary Last update: 17.01.2020

Responsible applicant and co-applicants

Employees

Project partner

Associated projects

Number Title Start Funding scheme
179876 New metrics for constraining multiple drivers of hazard and compound hazards 01.11.2018 Ambizione

Abstract

Large climate-related impacts on ecosystems or human societies are often related to multiple compounding weather and climate processes, which interact at different spatial and temporal scales. For instance, devastating floods are caused by specific spatiotemporal precipitation patterns in combination with antecedent soil moisture. Extremely low crop yields and vegetation activity are often related to multiple climate events occurring at different points in time, whose impacts compound each other. Identifying such compounding weather and climate conditions is very challenging due to the complexity of the involved systems as well as the very high dimensionality of multiscale climate and weather processes. The limited availability of observational data on impacts makes the development of new statistical approaches very difficult. Here, impact models such as vegetation models, crop models and hydrological models can serve as a tool to generate large amounts of realistic data. In this project, we will make use of a dynamic global vegetation model (LPX-Bern) to explore the ability of state-of-the-art machine learning techniques to identify climate and weather features that cause extreme reduction in vegetation productivity. LPX-Bern will be forced with very long simulations from a climate model to generate large amounts of impact data. We will then employ machine learning approaches such as Convolutional Neural Networks that allow an identification of the learned features to identify climate conditions that are associated with extremely low vegetation productivity through classification. We will assess the sensitivity of the classification to sample size and uncertainties in the impact variable, attributes that are important when transferring the approach to observational data. The resulting climate-feature impact relationships will be compared with output from other vegetation models.The highly interdisciplinary project will develop a novel approach to link extreme impacts with their multiple climatic drivers and will therefore provide new tools for compound event research. The comparison between climate-feature impact relationships across different models will help identify model uncertainties. The approach can also be used to assess climate-related risks in other sectors such as agriculture and flood protection. Ultimately, the project will pave the way to an improved assessment of climate risks.
-