Thesis subject

MSc thesis topic: Precision in Protection: Enhancing Regional Validation of Forest Change Maps using spatial modelling and machine learning

In an effort to combat deforestation-driven CO2 emissions, the European Union introduced the EU Deforestation Regulation (EUDR). This regulation requires companies to prove that their supply chains are deforestation-free, particularly for key import commodities such as soy, palm oil, wood, cocoa, coffee, rubber, and cattle.

Any product entering the EU market must be backed by geo-located evidence showing that it was not sourced from recently deforested land (After 2020). Any land cleared after this date cannot be linked to products sold in the EU.

Implementation of this regulation requires high-quality forest and forest change monitoring. In the absence of such high-quality regional forest monitoring, global forest and forest change products such as Global Forest Change by University of Maryland (Hansen et al., 2013) or Tropical Moist Forest forest cover change (Vancutsem et al., 2024) could provide a viable option.

However, accurately detecting fragmented and dynamic forests is challenging, introducing unwanted uncertainty in satellite-based forest change products. Additionally, these global products lack region-specific quality metrics, and their accuracy varies due to mapping challenges, such as limited cloud-free datasets and seasonal variations in dry forests. Studies have highlighted the over- and under-estimation of deforestation in different regions and years (Kinnebrew et al., 2022). Despite these limitations, no systematic approach exists to assess the regional quality of forest change products, limiting the use of openly accessible forest change products for policies targeting to combat deforestation.

Background

Location-specific quality assessment can be conducted by modeling spatial variation of map accuracy. Next to traditional methods like geographically weighted regression (See et al. 2014) and indicator kriging, new approaches leverage machine learning approaches such as spatial random forest (Talebi et al., 2022). In addition, advanced spatial interpolation techniques like Gaussian process modeling (Giorgi et al., 2023) or Bayesian kriging (Lezama-Ochoa et al., 2020) offer efficient ways to model location-specific map accuracy.

Recent advances in machine learning allow further improvement in modelling spatial relationships. A deep location-encoder model has been introduced to calculate semantic distances, helping to refine local accuracy assessments. Pre-trained location encoder models like SatCLIP [Klemmer et al., 2024] can compute semantic similarities of different locations, which can guide local accuracy assessments of maps.

The main goal of this study is to investigate efficient and scalable spatial prediction methods for location-specific quality assessment of forest change products. Depending on the specific research interests, two subtopics can be explored: (1) investigating advanced interpolation techniques such as Gaussian, Bayesian, or Random Forest interpolations or (2) exploring deep location-encoding models like SatCLIP [Klemmer et al., 2024] for assessing map accuracy of forest change maps. High quality, reference map is available for a region in Uganda as a test case. Other regions of study can be considered based on the availability of validation datasets.

Objectives

  • Assess the suitability of advanced spatial interpolation methods for location-specific quality of forest change maps.
  • Develop a local map validation model using similarities obtained from the pre-trained SatCLIP model.
  • Investigate and discuss similarities and differences between spatial interpolation methods and deep location encoding approaches.
  • Assess the local accuracy of forest change maps

Depending on the choice of the two topics, research objectives can vary.

Literature

Requirements

  • Deep learning or Spatial Modelling and Statistics
  • Geoscripting

Expected reading list before starting the thesis research

  • literature listed above

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