guest ::
| Home > Publications database > AI applications for the generation of a pan-European ecosystem reanalysis relying on the eLTER Research Infrastructure |
| Home > Publications database > AI applications for the generation of a pan-European ecosystem reanalysis relying on the eLTER Research Infrastructure |
| Poster (After Call) | FZJ-2024-05329 |
; ; ; ; ; ;
2024
This record in other databases:
Please use a persistent id in citations: doi:10.34734/FZJ-2024-05329
Abstract: Climate change, extreme weather events, and human activities increase the pressure and threats on ecosystem functions and services. Sustainable management of ecosystem functions and services is important, not only for preserving the ecosystems themselves, such as biodiversity, but also for human societies to provide essential resources like food, feed, fibre, wood, and clean water. eLTER RI (Integrated European Long-Term Ecosystem, critical zone and socio-ecological Research Infrastructure) will provide services based on aEurope-wide network of ecological stations with the aim to address these challenges. This will allow a wide range of stakeholders to analyse, monitor, and predict the state and evolution of ecosystems and to derive actionable knowledge.In this context, we aim to build a pan-European ecosystem reanalysis, which combines data from various sources (eLTER observations, remote sensing products, atmospheric reanalysis) with the Community Land Model, a state-of-the-art land surface model. This reanalysis will be developed at high spatial resolution (e.g., 3km), providing high-level data products on the states and fluxes of all European ecosystems and their evolution over time and complementing sparse observation data.Nevertheless, the generation of such an ecosystem reanalysis is associated with several obstacles, which we seek to solve with AI-based applications:(1) As the land surface and subsurface are characterised by high small-scale heterogeneity, large-scale high-resolution initial conditions are needed. This task is particularly challenging because the slow carbon pool needs a spin-up period of several centuries or even millennia. To address this, we will use AI to speed-up the computation of the spin-up and to generate the initial conditions by fusing different data sources (e.g., in-situ, remote sensing, and model outputs).(2) Additionally, data assimilation typically requires computation of an ensemble of dozens of members to yield reliable results, which is very intensive in time and computing resources, even more since transient runs are needed to account for slow processes like long-term changes in the carbon pools and subsurface water budget. In addition, the typical ensemble size used is still suboptimal and does not cover enough the state-parameter space. It is expected that a larger ensemble would yield better results. Combining data assimilation with AI surrogates allows for a fast processing of large ensembles. We will also use this hybrid approach for parameter estimation.(3) Furthermore, mechanistic models have practical limitations to further increase their spatial resolution. We will implement an AI application trained with in-situ and local remote sensing data to further downscale the modelled data products to a very high resolution (i.e., < 1km), eventually reaching relevant scales for decision support.Finally, despite being challenging, an expected co-benefit of these AI applications is their ability to estimate the uncertainties associated with, e.g., model-data fusion, downscaling, and parameter estimation, which will constitute a strong added value for the users of the ecosystem reanalysis data products.
|
The record appears in these collections: |
PDF
Fulltext