The PhD defence and trial lecture will be held in Auditorium 1 in The Geology Building. In some cases, it will be possible to attend the trial lecture and dissertation digitally, in that case a link to Zoom will be posted.
Trial lecture
Friday 11 October, 10:15-11:00, Aud 1, The Geology Building:
The past and future roles of machine learning in land surface modeling
Conferral summary (in Norwegian)
Planter er en vesentlig komponent i utallige sammenhenger i naturens kretsløp. For å forstå konsekvenser av klimaendringer, endringer i arealbruk og arealdekke, og tap av biodiversitet må vi ha robuste modellprediksjoner av vegetasjonsfordeling og -dynamikk, og tilbakekoblingene til klimasystemet. I sin avhandling har Lasse Torben Keetz brukt maskinlæring og andre datadrevne metoder for å videreutvikle to komplementære modelleringsverktøy, økosystemutbredelsesmodeller (EDM) og landoverflatemodeller (LSM) med fokus på hvordan de representerer vegetasjonsprosesser i dynamiske globale vegetasjonsmodeller (DGVM).
Main research findings
Popular scientific article about Keetz’s dissertation:
Unraveling the emerald planet: learning vegetation-environment interactions from data
Plants are vital components of many processes and interactions within the Earth system. Accordingly, improving our understanding of environmental pressures such as climate change, land use and land cover change, and biodiversity loss requires better predictions of vegetation distribution and dynamics and related Earth system feedbacks. Like many other domains, vegetation modeling has greatly benefited from the recent advent of novel machine learning and data assimilation algorithms paired with an unprecedented abundance of data, and from constantly improving hardware and software with easier accessibility. However, there is still great untapped potential for integrating available data and methods, which also hinges on improving technical infrastructure and data flow.
For his doctoral work, Lasse Torben Keetz and his co-authors thus aim to advance the understanding of vegetation-environment interactions from three data-driven directions. First, they used machine learning to predict the spatial distribution of Norwegian ecosystem types from available environmental raster data. Second, they developed new software designed to facilitate technical access to a Dynamic Global Vegetation Model (DGVM) embedded in a Land Surface Model (LSM), which are state-of-the-art approaches for simulating ecosystem-climate feedbacks on a global scale. Finally, they combined powerful data assimilation and machine learning methods to refine the DGVM/LSM’s parameters based on carbon and water exchange data, successfully improving agreements between model predictions and site-level observations.
Photo and other information:
Press photo: Lasse Torben Keetz, portrait; 900px. Photo: Private
Other photo material: Photo with description and credit as specified in the article above, size 1500px.