Learning to analyze and synthesize landscapes for ecological simulation

How to apply

Who to contact?	sinz@cs.uni-goettingen.de
Email subject	Start with "[ML Master]"
What to include?	All applications What you are studying (e.g. "Master in Applied Data Science") Statement of motivation (max 300 words) outlining your personal interest in our research areas. A short list of which relevant skills you have and which you would like to acquire. A list of courses you have taken and corresponding grades Known programming languages and experience Possible start date If you apply with your own project idea A concise (max 300 words) “research proposal” on a project you would be interested working on. This is not binding but helps to get the conversation started.

Background

Landscape composition and configuration have effects on the performance of species that may differ vastly on different scales and for different species. Plant or animal individuals experience landscapes locally. Thus landscape-based metrics are calculated locally using a weighted moving window to compute a metric from a local neighborhood for each cell of a rasterized landscape. However, currently implementations of this approach are computationally costly and not applicable to larger landscapes with billions of cells.

Randomly generated landscapes are used as inputs to ecological simulation experiments, for example, to study the impacts of land use on biodiversity. For such simulations, the landscapes need to fulfill certain sets of landscape measures but should vary in all other aspects. The challenge is now that (i) the measures are often correlated with each other and (ii) the parameters of landscape generators often not independent. For example, composition parameters (how much of the grid cells are habitat) impact configuration parameters (how habitat and non-habitat cells are distributed in space). Even with just two or three parameters, parameter sampling can be a real challenge. The current approach is very time-consuming to sample landscapes with a specific set of metrics from a large pool of randomly generated landscapes.

Project Description

The goal of this master thesis is to use deep learning techniques to make metric computation and random landscape generation more efficient. In the first part, you will train deep networks (e.g. a U-Net) to predict the metrics from a given landscape. The network will be trained on metrics on landscapes that have previously been computed. In the second part, this trained network will be used to generate landscapes with predefined metrics. Several approaches are possible, the most straightforward would be to “invert” the trained U-Net from the first part and optimize a landscape that matches a give output of metric values. This approach can also be extended to find landscapes that most closely match a given landscape but also the predefined metric values.