Student information

MSc thesis topic: Deep Learning AI Models for Crop Type Mapping with Sentinel-2

Our food supply utilizes few highly optimized crop species, to achieve high yields and compete in the global commodity markets (Van Zanten et al., 2014). Hence, intensive monocultures dominate the agricultural landscape contributing to biodiversity decline and soil degradation (Adegbeye et al, 2020; Mupepele et al., 2021). New ecological agronomic practices need to lift the environmental burden of food production and remote sensing can provide the necessary information, for instance, on crop rotations dynamics at a landscape level (Yang et al., 2020).

This Thesis will investigate crop rotation patterns and cover-cropping practices by classifying satellite image time series (SITS). Methodologically, it will use and design deep learning models to classify Sentinel-2 and PlanetScope data to identify crop type. Explainable-AI methods will then be applied to identify seeding and harvesting periods to identify fields where intermediate cover crops are cultivated. This research represents another step towards agriculture digitalization, highlighting the opportunities and limitations of remote sensing and data-driven approaches for designing more sustainable and diverse agricultural landscapes.

Relevance to research/projects at GRS or other groups

This Thesis is supervised by Marc Rußwurm who will provide expertise in deep time series classification models for crop type mapping and Paolo Dal Lago, whose research investigates ecological agricultural practices with remote sensing.

Research Questions & Objectives

  • Question 1: How can existing crop type mapping approaches be adapted to reveal the growing seasons?
  • Question 2: Can self-supervised learning be used for SITS segmentation and crop type labelling?
  • Explore the Microsoft Planetary Computer platform for data processing and analysis.
  • Objective 1 Implement and test deep crop type mapping approaches in the Netherlands.

Requirements

  • required: Remote Sensing & Deep Learning Courses
  • optional: Experience with Pytorch and Github

Expected reading list before starting the thesis research

  • Rußwurm, M., & Körner, M. (2020). Self-attention for raw optical satellite time series classification. ISPRS journal of photogrammetry and remote sensing, 169, 421-435.
  • Obadic, I., Roscher, R., Oliveira, D. A. B., & Zhu, X. X. (2022). Exploring self-attention for crop-type classification explainability. arXiv preprint arXiv:2210.13167.

Theme(s): Sensing & measuring, Integrated Land Monitoring