Tabea Rettelbach

Tabea Rettelbach

Facilitating Machine Learning on Super-High Resolution Earth Observation Data for Detecting and Quantifying Arctic Permafrost Thaw Dynamics

Permafrost, covering one quarter of Earth’s landmass, is undergoing significant change in a rapidly warming Arctic. Ground ice melt results in dramatic landscape reconfiguration in many regions due to soil volume loss, geomorphic change, and re-routing of hydrology. Permafrost soil carbon is mobilized and impacts the global carbon cycle by adding greenhouse gases to the atmosphere. In this PhD project, super-high resolution (<20cm) airborne stereo imagery and LiDAR point cloud data acquired with the AWI Polar-5 airplane in the Arctic will be processed and analyzed with machine learning methods to (1) map ground ice distribution and to (2) detect and quantify the presence and abundance of thaw-related landforms. Resulting data will be (3) used to train deep learning algorithms such as convolutional neural networks (CNN) to quantify such features in satellite imagery covering very large Arctic permafrost regions. Available datasets include tens of thousands of images captured with the innovative DLR MACS camera onboard Polar-5 in summer 2018 over Canada. Multi-year airborne LiDAR datasets exist for Alaska and Canada and additional data will be acquired in 2019-2021. The airborne data will be complemented by very high resolution (<50 cm) satellite imagery available at AWI. Their large spatial coverage results in raw datasets of tens of TB size that require processing and subsequent analysis with machine learning.

The project will result in a) the development of methods for handling huge very high resolution imagery datasets for Arctic landscapes, b) new machine learning algorithms to quantify ground ice and thaw- related features, c) the detection of early warning signs of rapid permafrost thaw in large regions of the Arctic using high resolution image and elevation datasets, and d) a prototypical implementation using current Data Science concepts for data integration, data cleaning, data analysis, data visualization, and metadata management. Potential platforms forming the basis for a scalable implementation environment in this project could be Apache Spark (UC Berkeley), Apache Flink (Berlin), or Apache Kafka. The experience gained in GeoMultiSens, one of the 10 BMBF-funded Big Data projects could help to advance rapidly in the current context.

Journal & Conference Publications

-

Oral & Poster Presentations at Conferences

-