Thesis subject

MSc thesis topic: Explainable land cover maps: deriving classes fromland surface traits

Traditionally, land cover maps are created using machine learning or deep learning techniques, which are largely black-box algorithms. Recent developments in vegetation trait mapping and radiative transfer modelling has resulted in an increasing availability of ready-made high processing level products that can help discern problematic classes, such as herbaceous from woody vegetation.

Given enough descriptive traits, next generation land cover products could ultimately be derived using an easily explainable decision tree or regression algorithm. Such a derived product would also allow the users to tweak class definitions to obtain customised land cover products for their specific usage needs.

Background

In this topic, auxiliary data to land cover that may improve land cover mapping results can be explored, such as leaf area index, vegetation continuous fields, sun-induced fluorescence and gross primary productivity, photosynthesis rate, canopy height, biomass, phenology, etc. There would be an opportunity to derive these traits manually using a radiative transfer model such as PROSAIL or SCOPE. An important component in this topic would be to understand what makes each land cover class distinct from others, so that useful data can be selected. And in the end, an attempt can be made at a land cover map that is defined as a combination of traits, compared with a traditional machine learning approach.

Relevance to research/projects at GRS or other groups

This research related to ongoing work on land cover mapping at the global scale, which is related to the Copernicus Global Land Services land cover maps, including the Land Cover and Forest Monitoring (LCFM) project.

Objectives and Research questions

  • What auxiliary data is available or could be produced that would allow us to better discern land cover classes of interest?
  • How important are features derived from each data source for land cover classification?
  • How well can land cover be (re)constructed from vegetation and non-vegetation traits, compared to traditional black-box methods?

Requirements

  • Required: Geoscripting, Remote Sensing
  • Optional: Machine learning, Advanced Earth Observation

Literature and information

  • Houborg, R., Fisher, J. B., & Skidmore, A. K. (2015). Advances in remote sensing of vegetation function and traits. International Journal of Applied Earth Observation and Geoinformation, 43, 1-6.
  • Quétier, F., Thébault, A., & Lavorel, S. (2007). Plant traits in a state and transition framework as markers of ecosystem response to land‐use change. Ecological monographs, 77(1), 33-52.
  • Rogers, C. A., & Chen, J. M. (2022). Land cover and latitude affect vegetation phenology determined from solar induced fluorescence across Ontario, Canada. International Journal of Applied Earth Observation and Geoinformation, 114, 103036.

Expected reading list before starting the thesis research

  • Van Bodegom, P. M., Douma, J. C., & Verheijen, L. M. (2014). A fully traits-based approach to modeling global vegetation distribution. Proceedings of the National Academy of Sciences, 111(38), 13733-13738.
  • Masiliūnas, D., Tsendbazar, N.-E., Herold, M., Lesiv, M., Buchhorn, M., & Verbesselt, J. (2021). Global land characterisation using land cover fractions at 100 m resolution. Remote Sensing of Environment, 259, 112409.

Theme(s): Modelling & visualisation; Integrated Land Monitoring