Project

Back to overview

Winter Climate Prediction with Machine Learning

Applicant Wegmann Martin
Number 195695
Funding scheme Spark
Research institution
Institution of higher education University of Berne - BE
Main discipline Climatology. Atmospherical Chemistry, Aeronomy
Start/End 01.12.2020 - 30.11.2021
Approved amount 99'406.00
Show all

Keywords (7)

prediction; variability; reanalysis; climate; machine learning; Seasonal Prediction; extreme events

Lay Summary (German)

Lead
Der globale Klimawandel erwärmt unsere Erde nicht nur, er führt auch dazu, dass unser Klima variabler und schwerer zu vorherzusagen wird. Vor allem im nordhemispherischen Winter kam es in den letzten Dekaden zu unvorhergesehen Klimaextrema und Wechselwirkungen im Erdsystem. Um zu wissen, welche Klimaextrema auch in Zukunft noch eintreffen können, hilft es, so weit wie möglich vergangenes Winterklima zu analysieren. Neue Klimadatensätze erlauben es der Menschheit, regionales und zeitlich hochaufgelöstes Klima der letzten 500 Jahre zu erforschen. Leider sind die Unsicherheiten für das Winterklima in diesen Datensätzen sehr viel höher als für das Sommerklima, da natürliche Klimaarchive hauptsächlich das Sommerklima aufzeichnen und instrumentale, meteorologische Aufzeichnungen so weit nicht zurückreichen.
Lay summary

Winterklimavorhersagen mit Hilfe von künstlicher Intelligenz

Unser Ziel ist es diese Informationslücken im Winterklima zu verkleinern. Im Detail werden wir dafür in einem innovativen Konzept zwei getrennte Felder der Klimawissenschaft zusammenbringen, die Paläoklimatologie und die saisonale Klimavorhersage. Dafür werden wir brandneue Methoden der Datenanalyse, im speziellen künstliche Intelligenz, einsetzen, welche uns erlauben werden aus den Informationen über das vergangene Sommerklima eine Vorhersage für das Winterklima zu errechnen. Diese Vorhersage kann dann dabei helfen, Lücken im Wissen über das vergangene Winterklima zu schliessen. 

 

Unsere Arbeit wird Forschern helfen neue, genauere Datensätze über vergangenes Klima zu erstellen und Entscheidungsträgern helfen Klimaextrema, welche wir heutzutage erleben, in einen besseren Kontext einzubetten. Ausserdem kann unsere Methode auch nach unserem Projekt dafür genutzt werden, tatsächliche, aktuelle saisonale Klimavorhersagen sehr günstig der Allgemeinheit zur Verfügung zu stellen. 

 

Direct link to Lay Summary Last update: 13.09.2020

Responsible applicant and co-applicants

Abstract

In-depth knowledge and understanding of past climate change and variability is crucial to create an adequate framework in which the current and future climatic change, climate extremes and their impacts (economic, social and ecological), can be assessed. Currently, reliable high-resolution climate information of the past 150 years is provided by climate reanalysis products. Essentially being climate models in tandem with weather observations, such data allow for analysis of past climate variability on high spatial and temporal resolution.New scientific endeavors (so called paleo-reanalyses) expand that timeline to the last 400-500 years by incorporating environmental proxy data and historical written evidence. Despite best efforts of assimilating a wide array of climate data, most such data are biased towards the boreal growing season, covering roughly the months April to September. Reliable large-scale information for the winter season is currently sparse to non-existent. Current winter climate over large parts of the Northern Hemisphere is changing rapidly and non-linearly (Winter 2019/2020 in Europe was the warmest on record). Attributing winter climate extremes to natural or forced (via greenhouse gas increase) variability is highly discussed in the scientific community, with no clear consensus emerging. Identifying a better framework for these climate winter extremes is therefore of both academic and societal importance, enabling better guidance on measures concerning future climates.A possible way to fill in missing winter climate information is the use of seasonal prediction products. Currently, potent seasonal prediction products are computationally extremely expensive. A new emerging field of seasonal prediction with low-cost machine learning algorithms shows promising results, is however untested for paleo climate applications.The aim of the WINTER project is therefore to combine cutting edge machine learning tools and long-term climate reanalysis products in order to improve the winter climate understanding for the past 500 years at hemispheric scale. More specifically, the project will use neural networks and random forest models to establish a gridded climate product for the boreal winter season that is more skillful than the baseline climate model winter performance, allowing for a better understanding of climate variability and extremes. Furthermore, the machine learning algorithm will be tested against state-of-the-art seasonal prediction products. As such, the project will also contribute to the improvement of prediction products in the seasonal-to-subseasonal (S2S) framework of the World Meteorological Organization (WMO).If successful, this project will create the most reliable high-resolution winter climate data for the past 500 years worldwide. The initial focus will be on the Northern Hemisphere, due to higher data availability, but can easily be adopted for the Southern Hemisphere. The combination of cutting-edge machine learning tools, novel paleo-reanalyses and modern seasonal prediction products is unprecedented and will contribute considerably to the understanding of seasonal climate prediction, climate reconstruction methods and large-scale climate variability.
-