Day 4

Detailed paper information

Back to list

Paper title Data Driven Analysis of Permafrost Region Disturbances - Spatio-temporal patterns, impacts and and key drivers
Authors
  1. Ingmar Nitze Alfred Wegener Institute - Helmholtz Center for Polar- and Marine Research Speaker
  2. Annett Bartsch b.geos GmbH, AT
  3. Sebastian Westermann University of Oslo, Norway
  4. Jaroslav Obu
  5. Alexandra Runge Alfred Wegener Institute Helmholtz Centre for Polar and Marine Research
  6. Konrad Heidler German Aerospace Center (DLR)
  7. Sophia Barth Alfred Wegener Institute Helmholtz Centre for Polar and Marine Research Potsdam/ University of Potsdam
  8. Anna Liljedahl
  9. Guido Grosse AWI (Alfred-Wegener-Institute Helmholtz Centre for Polar and Marine Research)
Form of presentation Poster
Topics
  • C1. AI and Data Analytics
    • C1.07 ML4Earth: Machine Learning for Earth Sciences
Abstract text In a rapidly warming Arctic, permafrost is increasingly affected by increasing temperatures and precipitation. Currently it is underlying around 14 Mkm² of the northern Hemispheric land mass and permafrost soils store about two times more carbon than the atmosphere. Thawing permafrost soils are therefore likely to become a significant source for carbon emissions under warming climate conditions. Gradual thaw of permafrost is well understood and included in Earth System Models. However, rapid or Permafrost Region Disturbances (PRD) such as wildfires, retrogressive thaw slumps or rapid lake dynamics are widespread across the Arctic permafrost region. Due to a combination of scarce data and rapid dynamics, with process durations from hours (e.g. wildfire, lake drainage) to years (lake expansion), there is still a massive lack of knowledge about their distribution in space in time. In the rapidly warming and wetting climate they are potentially accelerating in abundance and velocity with significant implications to local and global biogeochemical cycles as well as human livelihoods in northern high latitudes. Despite their significance, these disturbances are still not thoroughly quantified in space and time and thus accounted for in Earth System Models due to the past and current lack of quantification.
Historically, remote sensing and data analysis of Arctic permafrost landscape dynamics was highly limited by data availability. The explosively expanding availability of remote sensing data over the past decade, fuelled by new satellite constellations and open data policies, opened up new opportunities for spatio-temporal high resolution analysis of PRD for the research community. This data abundance, in combination with new processing techniques (cloud computing, machine learning, deep learning, unprecedented fast data processing), led to the emergence and publication of new, publicly and freely available datasets. Such datasets include permafrost-related model-based panarctic datasets (e.g ESA Permafrost CCI Ground Temperature, Active Layer Thickness), machine- and deep-learning based remote sensing-based datasets (e.g ESA GlobPermafrost Lake Changes, Retrogressive Thaw Slumps, ArcticDEM), and synthesis data from different sources (e.g. the Boreal-Arctic Wetland and Lake Database BAWLD).
Combining these rich datasets in a data science approach and leveraging machine-learning techniques has the potential to create synergies and to create new knowledge on the spatio-temporal patterns, impacts, and key drivers of PRD. Within the framework of the ESA CCI+ Permafrost and NSF Permafrost Discovery Gateway Projects, we apply a synthesis of publicly available permafrost-related datasets of permafrost ground conditions (ALT, GT), climate reanalysis data (ERA 5), and readily available or experimental remote sensing-based datasets of permafrost region disturbances.
We will (1) analyze spatio-temporal patterns, correlations, and interconnections between different parameters, (2) retrieve the importance of potential input factors (climate, stratigraphy, permafrost) on triggering RTS using machine-learning methods (e.g. Random Forest Feature Importance) and also experimenting with more advanced deep learning methods such as LSTM to retrieve temporal inter-connections and dependencies.First analyses of the spatial patterns of lake dynamics on continental scales (Nitze et al., 2018, > 600k individual lakes) reveal enhanced lake dynamics in warm permafrost close to 0 °C. Furthermore, we found enhanced ALT thickness variability in burned sites.
With analyzing and inferring key influencing factors, we may be able to predict/model the occurrence and dynamics of permafrost region disturbances under different warming scenarios. As PRD’s are still not sufficiently accounted for in global climate models, this and follow-up analyses could help fill a significant knowledge gap in permafrost and climate research.