PhD project offered by the IMPRS-gBGC in Jan 2024

Harnessing computer vision and machine learning for a renewed cartography of human impact on landscapes

Joachim Denzler , Gregory Duveiller , Hanna Meyer , Markus Reichstein

Project description

We live in an unprecedented period where geospatial information and remote sensing imagery is becoming ubiquitous and increasingly detailed, making it very easy to make maps of land cover. However, actually mapping the impact and intensity of human actions on the landscape in a consistent way is far from straightforward, as the processes behind are complex, multi-faceted and geographically dependent. And yet there is currently a very strong demand for such a map coming from various sectors: (1) Such a map would be an invaluable proxy for the biodiversity community to link with in-situ databases of species occurrence (GBIF, iNaturalist, FloraIncognita) and thereby progress towards mapping essential biodiversity variables; (2) Restoring degraded landscapes is a major target of the times (2021-2030 is the UN Decade on Ecosystem Restoration), but this often involves establishing agro-ecosystems with sustainable land management practices (e.g. agroforestry) that typically have more complex spatial patterns, which need to be identified, characterized and mapped for proper monitoring towards restoration goals; (3) Properly accounting for the sources and sinks of carbon from the land sector requires a distinction of natural versus managed lands, which currently can only be done at very course scale with crude assumptions.
The PhD will aim to apply novel AI and computer vision techniques to high and very high spatial resolution imagery to attempt to characterize and map the human impact on the landscape. As increased land management can both lead to an increase in homogeneity or evenness on one side or increasing spatial diversity on the other, the objective will be to derive two main axes of variation: the first ranging from intact landscapes to extremely anthropic landscapes, while the second reflecting the degree of heterogeneity of the landscape. A key methodological problem that must be addressed is that these land use dynamics occur across non-stationary environments, in which machine learning models will suffer from training bias [1]. The PhD will explore concepts from federated learning and test them on specific land characterization tasks to infer stable features [2]. Stable features show stable contributions over different, diverging environments and are assumed not to carry any bias generated from the local data distributions. The hypothesis is that machine learning models are more robust against distribution shifts in the data if stable features are estimated that work well even under changing environments and that can be combined with environmental-specific features that consider local situations. Links will be addressed to causally informed machine learning models and hybrid modeling [3].

Working group

The PhD candidate will work in the Department of Biogeochemical Integration at the Max Planck Institute for Biogeochemistry and is expected to graduate at the Friedrich Schiller University, Jena, in collaboration with the Remote Sensing and Spatial Modelling Research Group at the University of Münster. The working group covers long-standing expertise in the overall topic, data streams and methods, including: Remote sensing (Gregory Duveiller, Hanna Meyer), landscape ecology (Hanna Meyer, Gregory Duveiller, Markus Reichstein), computer vision and machine learning (Joachim Denzler, Markus Reichstein), spatio-temporal modelling (Hanna Meyer, Gregory Duveiller) and ecosystem functioning (Gregory Duveiller, Markus Reichstein). The PhD candidate will be also associated with the ELLIS Unit Jena.

Requirements for the PhD project are

Applications to the IMPRS-gBGC are open to well-motivated and highly-qualified students from all countries. Prerequisites for this PhD project are:

either Master in Computer Science or related areas with a minor/experience in Earth System Science, Agronomy, Forestry, Geography or Ecology, or a Master in Earth System Science or related areas (Agronomy, Geography, Forestry or Ecology) with a minor/experience in computer science
programming experience in a modern computing language (Python, R, Julia)
background in machine learning
experience with deep learning models is beneficial but not mandatory
interest in sustainability and sustainable land management
interest in working in an interdisciplinary team of ecologists and computer scientists
very good oral and written communication skills in English

The Max Planck Society (MPS) strives for gender equality and diversity. The MPS aims to increase the proportion of women in areas where they are underrepresented. Women are therefore explicitly encouraged to apply. We welcome applications from all fields. The Max Planck Society has set itself the goal of employing more severely disabled people. Applications from severely disabled persons are expressly encouraged.

References

[1] Ludwig M, Moreno-Martinez A, Hölzel N, Pebesma E, Meyer H: Assessing and improving the transferability of current global spatial prediction models. Global Ecology and Biogeography 00: pp. 1–13. 2023.
[2] Cui P, Athey S.: Stable learning establishes some common ground between causal inference and machine learning. Nature Machine Intelligence; 4: pp. 110-115. 2022.
[3] Reichstein M, Camps-Valls G, Stevens B, Jung M, Denzler J, Carvalhais N, Prabhat: Deep learning and process understanding for data-driven Earth system science. Nature. 566 (7743) : pp. 195-204. 2019.