Bayesian Deep Learning for crop classification – measuring uncertainty in Remote Sensing predictions
To map tropical crops like oil palms or cocoa, uncertainty in the model's prediction must be quantified and considered for actual sustainable growth, production and supply chains. Yet, map uncertainty is far from being standard in applied deep learning systems. In this thesis, we want to explore the applicability of current methods to crop classification and segmentation in order to produce uncertainty maps alongside standard classification maps. The goal of this thesis is to implement existing Bayesian deep learning frameworks to investigate their applicability to crop classification in Remote Sensing. As a first step, the student would need to get a coarse overview of current state-of-the-art methods in Bayesian deep learning and evaluate their applicability to crop classification. Then, the most promising approach can be implemented in our existing deep learning pipeline with input from satellite imagery and different crops as ground truth. We would start with a binary segmentation task that can be easily extended to multiclass tasks. Finally, the student would evaluate the uncertainty measures both qualitatively and quantitatively. Prerequisites are a strong interest in deep learning as well as good programming skills in Python and common machine learning libraries e.g. PyTorch.
Contact: Prof. Jan Dirk Wegner
Physically based, deep learning forecasting model for urban floods
This project aims to develop a forecasting model for urban floods which integrates physical modelling with data-driven deep learning. We aim to develop a forecasting model for urban floods, by tightly integrating physical modelling with data-driven deep learning. Our primary application scenario are urban pluvial floods that occur when precipitation cannot be fully absorbed by the drainage system, therefore causing flooding and substantial damages, as well as disruption to socio-economic activities. The fast occurrence and relatively short duration of urban floods (“flash floods”) means that physically-based models are of limited use, due to their long run times. In this project, we will develop new Machine Learning (ML) methods to generate flood predictions with sufficient lead time, such that they can be used to alert the population and to plan mitigation and rescue actions. Our hypothesis is that by exploiting hydraulic modelling knowledge, deep learning need not “ statistically learn physics from scratch”. By tightly integrating the two modelling approaches, we aim to get the best of both worlds: interpretability, adherence to physical constraints and the predictive power and speed of neural networks. The input data for the ML flood forecasting model will be rainfall forecasts provided by meteorological services (e.g. from weather radar) and digital surface models. The ML-based flood model will produce spatially explicit, two-dimensional flood hazard maps with water depth, flood extent and flow velocity information. The goal of this master thesis project is to integrate physical models with deep learning for flood forecasting using 3D city models, precipitation nowcasting data and simulation data from a physically-based model. Prerequisites are a strong interest in deep learning as well as good programming skills in Python and common machine learning libraries e.g. PyTorch.
Contact: Prof. Jan Dirk Wegner
Tree instance segmentation in UAV-LS point clouds
We will develop a deep learning method for tree instance segmentation in point clouds of forests acquired with a laser scanner mounted on a UAV. Laser scanners suitable for unmanned aerial vehicles (UAV-LS) allow collecting finely detailed three-dimensional (3D) remote sensing data over forested areas. The density of resulting point clouds allows to resolve single trees and the different parts of each tree (i.e., stems, branches, and leaves), ultimately allowing the measurement of biomass and biomass dynamics of each tree. The use of UAV-LS data for deriving single-tree measurements is a rapidly increasing field of research with potential to open new frontiers in the field of fully-airborne forest inventories. The first step required to enable the direct measurement of tree properties is the detection of single trees and their parsing into different components (i.e. stem, live, and dead branches). Current methods are typically developed on small datasets and rely on rather lengthy pipelines characterized by the need to manually input several parameters. To overcome such manually intensive and rarely transferrable methods, the aim of this master thesis is to develop a machine learning/deep learning approach to segment individual trees and to parse them into different components like stem, live and dead branches. For this purpose, we have manually labelled the several thousand trees in point clouds covering different forest types. This Master thesis project is a collaboration of the Norwegian Institute of Bioeconomy Research (NIBIO) and ETH Zurich. Eligible candidates should have a strong background in deep learning and python programming. Experience with point cloud processing is a plus.
Contact: Prof. Jan Dirk Wegner
Planet formation and evolution, planetary interiors, extra-solar planets
Giant planet formation and evolution
Giant planets play a key role in the formation and dynamical evolution of planetary systems. Their origin, however, is still poorly understood. In this project we will use numerical simulations to model the formation and evolution of giant planets with the aim to provide predictions on their compositions and final masses of the planets. We will explore the formation of the outer planets in the solar system as well as of giant exoplanets.
Contact: Ravit Helled
Planets orbiting other stars are detected at a rapid pace. When both the mass and the radius of a planet are measured, the average density of the planet can be determined. In order to understand what exoplanets are made of, structure models are required. In this project we will characterise exoplanets using interior models, to put limits on their bulk compositions and internal structure, and when possible, link it to their formation and evolution histories.
Contact: Ravit Helled
Giant Impacts of clumps
Protoplanetary disks that are massive enough can fragment leading to the formation of clumps that could evolve to become giant planets. Simulations suggest that several clumps form in such disks. The interactions between these clumps and their potential collisions that can lead to both mergers and fragmentation are poorly known. In this project we plan to simulate the collisions between gaseous clumps using SPH simulations and determine their fate for a large range of initial and impact conditions.
Contact: Ravit Helled & Joachim Stadel
Theoretical Astrophysics and Cosmology
Signature of modified gravity in cosmological observables
Einstein’s general relativity works well in the solar system, but the problems in the standard model of cosmology may indicate the breakdown of general relativity on cosmological scales. Compared to the standard model, we compute the predictions in modified gravity theories in cosmological observables such as galaxy clustering, weak lensing, and cosmic microwave background anisotropies. We then quantify what is needed in future large-scale surveys to rule out or confirm such modified gravity theories. Requirement: good understanding of cosmological observables and general relativity
Gravitational waves from nonlinearity
Observations of gravitational waves on cosmological scales would provide a smoking gun evidence for inflationary expansion in the early Universe. Future surveys aim to reduce the upper limit on the amplitude of primordial gravitational waves by an order of magnitude. However, nonlinear evolution in density fluctuations in the late time can produce small corrections to the gravitational wave amplitude. We compute the level of this contribution and investigate how they can be separated from the primordial gravitational waves. Requirement: good understanding of 2nd-order relativistic perturbation theory
Contact: Prof. Jaiyul Yoo
Large-scale structure of the Universe, relativistic N-body simulations
My group works with N-body simulations of cosmic structure formation, ray tracing and statistical analysis of simulation data. We can offer projects with the aim to extract statistical features from the distribution of matter in the Universe. Examples include
- A project to develop improvements to the two-point correlation function estimator for large, high-quality data sets
- A project to study the effect of relativistic frame-dragging on the statistics of gravitational weak-lensing tangential shear
Contact: Prof. Julian Adamek