Student information

MSc thesis topic: Assessing spatial fairness of maps with deep learning

The quality of remote sensing based land monitoring is often varies across space. In data rich regions such as Europe and North America, the maps tend to have high quality. On the other hand, dry sub-tropical and sahel regions map are mapped with low quality. Assessing the local variations in map quality can reveal issues connected to geographic bias and spatial fairness. This information is valuable to inform which map to use where for users and guides future data collection and annotation efforts towards areas identified as low accuracy-regions for map producers.

Spatial variation of map accuracy can be modelled using different techniques such as indicator kriging (Tsendbazar et al 2015) and geographically weighted regression (GWR) (See et al 2014). However, these methods are known to be time consuming and require a large number of reference sample sites with a good geographical spread.

There have been new developments in machine learning to model spatial relationships in different variables of interest. Semantic distance calculation through deep location encoder model has been recently introduced. Pre-trained location encoder models like SatCLIP [Klemmer et al., 2024] can be used to calculate spatial similarities of different locations which guides local accuracy assessments of maps.

This study aims to investigate the effectiveness of deep location encoding models like SatCLIP [Klemmer et al., 2024] for assessing local map accuracy and spatial fairness of land use and land cover maps. Depending on the interest, the region of study can be defined based on the availability of validation datasets.

Objectives

  • Develop local map validation model using similarities obtained from the pretrained SatCLIP model.
  • Investigate and discuss similarities and differences between classic geo-statistical methods and deep location encoding approaches.
  • Assess spatial fairness of LULC maps by creating a map of local land cover accuracy

Exected reading list before starting the thesis research

  • Konstantin Klemmer, Esther Rolf, Caleb Robinson, Lester Mackey, and Marc Rußwurm. Satclip: Global, general-purpose location embeddings with satellite imagery. arXiv preprint arXiv:2311.17179, 2023. In particular, see source code https://github.com/microsoft/satclip and associated notebooks!
  • Tsendbazar, N.E., de Bruin, S., Fritz, S., & Herold, M. (2015). Spatial Accuracy Assessment and Integration of Global Land Cover Datasets. Remote Sensing, 7, 15804-15821
  • Stefanos Georganos, Tais Grippa, Assane Niang Gadiaga, Catherine Linard, Moritz Lennert, Sabine Vanhuysse, Nicholus Mboga, Eléonore Wolff & Stamatis Kalogirou (2021) Geographical random forests: a spatial extension of the random forest algorithm to address spatial heterogeneity in remote sensing and population modelling, Geocarto International, 36:2, 121-136, DOI: 10.1080/10106049.2019.1595177
  • Møller, A.B., Beucher, A.M., Pouladi, N., & Greve, M.H. (2020). Oblique geographic coordinates as covariates for digital soil mapping. SOIL, 6, 269-289

Requirements

  • Deep learning
  • Geoscripting

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