Research and Applications
The main research focus of ML4Earth is to bring together novel AI methods for Earth observation purposes and Earth and Climate Sciences. We are covering six major research branches, each linked to a specific application field.
Physics-aware
machine learning
With physics-aware machine learning, we use physical laws and physics-domain-specific knowledge to improve the performance and validity of our data-driven machine learning models.
Application: Hydrology (in collaboration with the University of Bonn)
While hydrological models do exist already, Climate Change is fueling the need for more precise and reliable forecasts of droughts or flood events. To do so, we dedicate our research to the development of novel methods that combine physics-aware machine learning with Earth observation data.
With reasoning, we go beyond perception and the direct recognition of materials, objects, and phenomena, or on measuring geo-physical parameters with Earth observation. We aim at higher-level semantic and process understanding by exploring the spatial relationships between entities and the temporal behavior of objects, as well as vision-language reasoning.
Application: Rapid permafrost thawing (Partner: Alfred Wegener Institute)
The much-accelerated thawing of permafrost areas in the arctic goes hand in hand with drastic threats such as coast erosion or the excess emission of methane. Due to Climate Change, this phenomenon is affecting increasingly large areas that can only be monitored from space. The challenge, however, is to detect permafrost thawing on satellite images in the first place and do so in an automated way for the large areas to be covered. Automated machine-learning algorithms would be a great solution to this problem. Existing algorithms, however, still largely depend on human annotations of the data to detect the delicate signatures of permafrost thawing in the huge and diverse amount of data. We facilitate spatio-temporal correlations in the related processes to reduce human annotations to a minimum, and to be able to apply our trained models for the arctic as a whole.
Uncertainty Quantification
The large-scale physical or bio-chemical quantities retrieved from Earth observation data are often used in data assimilation and in decision making, for example the Green Deal and the UN Sustainable Development Goals (SDGs). Therefore, besides high accuracy, traceability and reproducibility of results, quantifying the uncertainty of these predictions from an AI algorithm is indispensable towards a quality and reliable AI4EO. We develop methods that can effectively quantify well calibrated data- and model-uncertainties and detect out of distribution scenarios.
Application: Sea Level Budget (Partner: University of Bristol)
One of the most significant consequences of Climate Change is sea level rise. Before making precise predictions, however, one needs to understand the various interlinking processes that do affect the sea level and in particular related uncertainties. We are therefore dedicating this research branch to quantifying uncertainties of AI models, which are being used to assess the sea level budget under consideration of the complex set of contributing parameters, for example via monitoring glacial changes with Earth observation satellites.
Explainable AI
In very simple terms, explainable AI (or xAI) is a research field that aims at making it understandable and interpretable for the user of AI algorithms, i.e., explaining how the algorithm is approaching a solution.
Application: Extreme weather events (Partner: University of Leipzig)
The IPCC reached great consensus that extreme weather events will become more likely in the future due to Climate Change. For example, extreme droughts affect the future of existing forests, while extreme precipitation events can quickly turn into a threat for residential areas. Each of these extreme events alone has only limited explanatory power for making reliable predictions. With AI models, however, we can study the long-term dynamics of extreme events themselves and their occurrence, as well as correlated parameters. With xAI, in particular, we aim at gaining access to the complex underlying processes needed to not only make predictions, but to take effective action.
Sparse Labels & Transferability
When training AI classification models on Earth observation images, reference labels from the ground, so-called ground-truth labels, act as reference information for the algorithm. Such labels may provide very detailed information on a particular area, for example, which parts of the image are covered by crop fields, forests, or urban settlements. When applying a classification model that has been trained with rather localized information to different cultural zones or geographical areas, the classifier may likely be less reliable. To counteract, we dedicate this research branch to making our models transferable.
A second issue is that acquiring reference labels for Earth observation data is a highly time-consuming task for human annotators. As a result, AI models often have to be trained with sparse label sets. To do so most efficiently is subject of our research.
Application: Soil parameter mapping (Partner: DLR)
We focus on the need of public authorities to map physical and chemical soil parameters at large scales and over time. Here, the mapping is challenged by the fact that precise reference labels via soil samples cannot be provided at large scales. Also, for deriving soil parameters via Earth observation satellites, high-resolution spectral information are needed that only a few satellites can provide. The recently launched and most promising satellite mission EnMAP, however, is not designed to monitor large scales on a regular basis. This is where our research on sparse labels and transferable models comes into play.
Complex Structures
In order to model geo-processes, one requires climate and Earth-System variables that are very heterogeneous in scale, quality, and dimensions. Such complex data sources can be three-dimensional spatial information on the land surface gained via radar remote sensing, the temporal evolution of the land surface, or other data streams from sensors on ground. Analyzing these complex geo-spatial-temporal data streams using deep learning is subject to this research branch. In particular, we aim for using graph convolutional neural networks and understanding the involved causal relations.
Application: Tipping points (In collaboration with TUM Professorship for Earth System Modelling)
Tipping points can be defined by abrupt and irreversible changes in the climate system. Their nature and the underlying processes, however, are still subject to discussion. We develop novel Deep Learning models to analyze multi-dimensional data streams and unveil the complex processes of our Earth’s climate system.