(No documents in this view)
Modelling, Prediction and Anomaly detection of Earth Surface Dynamics
A TUM-ICL Joint Academy of Doctoral Studies (JADS) Project
Geo-physical processes on Earth are governed primarily by complex, intertwined, and often non-stationary dynamical systems on various spatio-temporal scales. While it is possible to identify some isolated subprocesses, many of them interact with each other according to inscrutable and opaque causal patterns. Studying such processes by developing process-based models requires expert knowledge, but is limited by the availability of large and reliable data stocks, usually acquired in-situ at irregular spatial and temporal frequencies. With the launch and reliable operation of modern Earth observation satellite missions, it now became possible to monitor the Earth at various spatial and temporal scales with different sensor modalities. This coincides with the advent of powerful data-driven modeling techniques provided by the machine learning community and advances in stochastic modeling for efficient analysis of sequential measurements. If this expertise can be joined, together with modern big geospatial data management solutions, it seems possible to gain deeper and more detailed insights into the Earth system.
We identified two sub-processes to investigate along the lines of these ideas, i.e., the seasonal growth of crops cultivated on the Earth's surface and the currents occurring in the oceans. Although occurring at different spatio-temporal scales, we believe they can be modeled in a generic data-driven way. To demonstrate the generalizability of our expected findings, we will apply them to model atmospheric processes assumed to drive the two originally considered systems.
The overall goal of the project is to model dynamic processes occurring on and within the Earth system at different spheres and to develop new methodology and software for their prediction, scenario generation, anomaly detection and uncertainty quantification. The modelling will be informed by dense (temporal, spatial and spectral) sequences of Earth observation satellite data (available from TUM and its associated partner institute at DLR). The project combines expertise from machine learning and remote sensing (TUM) with stochastic analysis, statistics and data assimilation (Imperial).
The dynamics of the Earth system are very diverse and we will develop new methodology based on three application areas:
- Lithosphere (Land surface): The goal is to model and develop models to describe the growth of cultivated crops and to derive crop yield predictions based on phenology & climate variables.
- Hydrosphere (Oceans): We aim to model and predict upper ocean transport of acidity, chemical concentration, heat and salinity.
- Atmosphere: We will focus on atmospheric wind field modelling and prediction.
The project will approach its key objectives following a combined bottom-up and top-down strategy: On the one-hand, we will be using data-driven techniques from remote sensing and machine learning to evaluate, classify and explain high-resolution Earth observation satellite data (TUM expertise). The results of this analysis will feed into the development of a more coarse-grained stochastic model for dynamical systems, which typically depend on high-dimensional state vectors. The stochastic models will be based on insights from physical/biological models and recent advances in spatio-temporal statistics and data assimilation (Imperial expertise).
The machine learning models used for addressing the questions related to remote sensing and Earth observation will comprise, i.a.,
- deep feed-forward neural networks,
- recurrent neural networks,
- hybrid modeling schemes, and
- spatio-temporal attention-based architectures.
Through explicitly modeling the uncertainty within the model and the data, we will also develop tools for anomaly detection.
We will combine a variety of data sources for our project, including satellite remote sensing and in-situ observations. For example, for monitoring crops, optical and radar data is available from Copernicus’s Sentinel satellites. The atmosphere is observed by Aeolus (wind field currents and directions) and MetOp (atmospheric parameters, e.g., temperature, trace gases, etc.). Upper ocean dynamics are measured
through drifter (NOAA) and satellites (Suomi NPP and the upcoming SWOT mission). Continuous and prioritized access to Earth observation satellite data is ensured via TUMs partner institutes at DLR.
TUM and Imperial College London (ICL) have recently combined their strengths in research, innovation and education in a strategic flagship partnership. Within the context of this bi-lateral cross-disciplinary program, a cohort of six joint projects is funded in a field of interdisciplinary cutting-edge research, in which both universities enjoy an outstanding reputation.
Projects will be supported through their individual institutions based on doctoral training stipends and relevant mobilities and consumables funds for a total of four years.
For further information, please refer to the JADS website.
- TUM Graduate School: The Joint Academy of Doctoral Studies welcomes six projects for the 2021 cohort (12.07.2021)