Student information

MSc thesis topic: Deforestation Prediction with Spatio-Temporal Deep Learning Models

Tropical deforestation remains a major global challenge, threatening biodiversity, carbon storage, and the livelihoods of local communities. Accurate and early prediction of deforestation events is vital for implementing timely conservation strategies.

Recent advances in deep learning and remote sensing have improved predictive models based on environmental variables [1][2]; however, most approaches treat different time periods as independent, failing to leverage the full temporal dynamics of deforestation. Although a few studies have started to explore spatio-temporal deep learning models for deforestation forecasting—such as the use of deep convolutional networks [3]—the field remains largely underexplored. In particular, Vision Transformers (ViTs), a powerful class of models for capturing complex spatio-temporal dependencies, have not yet been applied to the problem of deforestation prediction. This thesis aims to close this gap by developing and comparing spatio-temporal models (3D FCN, ConvLSTM, and ViT) against a 2D CNN baseline, for binary classification of deforestation risk using environmental variables and satellite alert systems (GLAD [4] and RADD [5]) as reference data.

Background

The student undertaking this project will be expected to train and validate AI models, specifically 3D FCN, ConvLSTM, and ViT, to predict deforestation using environmental variables. A strong knowledge of AI and machine learning workflows is recommended for this project.

Software: Pytorch (python) ArcGIS;QGIS

Objectives

  • Conduct a literature review on spatio-temporal deep learning approaches in environmental sciences.
  • Develop and implement spatio-temporal models (3D FCN, ConvLSTM, and ViT) for binary classification of deforestation risk.
  • Integrate historical sequences of environmental variables into model training.
  • Apply the methodology in three target regions: Gabon, Kalimantan, and Brazil.

Requirements

  • Geo-scripting course
  • Advanced Earth Observation
  • Machine Learning

Literature

  1. Mayfield, Helen J., et al. "Considerations for selecting a machine learning technique for predicting deforestation." Environmental Modelling & Software 131 (2020): 104741.
  2. La Rosa, L.E.C., Ferrari, F., Bezerra, F.G.S., de Souza, R.A., Maretto, R.V., de Aguiar, A.P.D., Vinhas, L., Happ, P.N., Feitosa, R.Q., 2024. Xgboost and multitemporal deter data for deforestation forecasting. ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences 10, 193–198.
  3. Ball, James GC, et al. "Using deep convolutional neural networks to forecast spatial patterns of Amazonian deforestation." Methods in ecology and evolution 13.11 (2022): 2622-2634.
  4. Hansen, Matthew C., et al. "Humid tropical forest disturbance alerts using Landsat data." Environmental Research Letters 11.3 (2016): 034008.
  5. Reiche J, Mullissa A, Slagter B, Gou Y, Tsendbazar N, Odongo-Braun C, Vollrath A, Weisse M, Stolle F, Pickens A, Donchyts G, Clinton N, Gorelick N & Herold M, (2021), Forest disturbance alerts for the Congo Basin using Sentinel-1, Environmental Research Letters.

Theme(s): Sensing & measuring; Modelling & visualisation