Research

I am dedicated to use Machine Learning for Good. Since our planet needs urgent climate action, I am using my skill to advance and accelerate climate applications. Here you can find some of my research projects.

Snowdragon: Snow Stratigraphy Classification

An early classification of snowflakes by Isreal Perkins Warren in "Snowflakes: A Chapter from the Book of Nature", 1863.

The cryosphere plays a crucial role in stabilizing earth’s climate. For climate change adaption it is essential to know in which state the cryosphere is in order to analyze e.g. polar tipping points. One important tool to discern snow and climate mechanisms is snow classification and segmentation. This can be done with a Snow Micro Pen (SMP) in a fast and high-resolution fashion. However, the resulting signal profiles could so far only be labelled manually. The manual labeling takes a lot of time, practice and becomes overall infeasible for large datasets. We showed that ML algorithms - among others an LSTM - can be used to segment and classify SMP profiles automatically. We provided a comparison of 14 models to enable practicioners to choose a suitable model for their specific task and dataset. Through this study, it became possible to analyze the MOSAiC SMP dataset - a unique and large dataset that makes longterm analysis of snow on Arctic ice possible for the first time. This shows, that snowdragon will help scientists to analyze the effects of climate change on the cryosphere faster and more efficiently.

For more details:

Caiclone: Extratropical Cyclone Prediction

Extratropcical Cyclone over the Hudson Bay captured by Suomi NPP in 2016, stored in NASA's Visible Earth catalogue

Extratropical cyclones influence and drive our global climate system. Extratropical cyclones can produce extreme weather events, causing flooding and storms at some locations and droughts at others. Climate models still struggle with simulating extratropical cyclones and especially their regional effects for different climate scenarios. For climate adaptation, it is however, crucial to know how the intensity of extratropical cyclone will develop for different temperature gradients. Machine learning methods can help in this process to accelerate and advance extratropical cyclone prediction. As a first step towards this goal, we are currently creating a dataset consisting of temperature gradient precursors and cyclone labels. The dataset uses cyclone tracks recovered from simulations by Hodges tracking algorithm. The “caiclone” dataset will make it possible to evaluate different machine learning models on extratropical cyclone image classification tasks. As baselines we propose CNN and an adapted Fourier Neural Operator architecture.

More details will follow soon.

Iceagle: Ice Lead Analysis with GNNs

Icelead Photograph by Joe MacGregor, stored at NASA's Airborne Science Program.

Ice lead analysis is an essential task for evaluating climate change processes in the Arctic. Ice leads are narrow cracks in the sea-ice, which build a complex network. Dynamics of ice leads over time remain largely hidden and unexplored until today. We propose to to interpret ice leads as more than just airborne images, but as dynamic networks. This new network perspective on ice leads could be of great interest for the cryospheric science community since it opens the door to new methods for ice lead forecasting and tracking. We found out that current network analysis methods cannot be easily used for ice lead networks, since they are fundamentally different from common networks. Hence, our work is also a call for extending existing network analysis toolkits to include a new class of real-world dynamic networks.

For more details: