Internship: Cocoa Farm Mapping with Remote Sensing, ML & RAG

Cocoa is one of the most important agricultural commodities in West Africa, with Côte d’Ivoire producing more than 40% of the world’s supply. Accurate mapping of cocoa farms is essential for improving supply-chain transparency, supporting zero-deforestation commitments, and ensuring sustainable sourcing. However, cocoa landscapes are highly heterogeneous, complex, and often difficult to distinguish from other tree crops using traditional machine learning approaches. The need for high-resolution spatially explicit cocoa maps that generalize across ecological zones has never been greater.

Remote sensing has long supported commodity-crop monitoring, but conventional methods such as random forests or standard deep learning classifiers require large amounts of labelled data and often struggle to capture fine-scale structural variations in cocoa agroforestry systems. Recent advances in retrieval-augmented generation (RAG) offer a promising new direction. RAG enables models to dynamically retrieve relevant training examples, auxiliary information, or contextual embeddings during inference, improving classification accuracy and generalization, particularly when class boundaries are ambiguous.

In this internship, the student will explore how RAG-enhanced machine learning can improve cocoa farm mapping compared to existing traditional approaches. The student will have access to a large, proprietary database of georeferenced cocoa farms in Côte d’Ivoire provided by ofi, enabling the development, evaluation, and validation of next-generation cocoa-mapping models.

This internship will address the following core questions:

1. How can retrieval-augmented generation (RAG) improve the classification and mapping of cocoa farms using satellite remote sensing?
2. How do RAG-enhanced models compare with traditional machine learning approaches (e.g., Random Forest, CNNs, Transformers) when applied to heterogeneous cocoa agroforestry landscapes in Côte d’Ivoire?
3. Does integrating retrieved contextual information, reference parcels, or learned embeddings improve generalization across ecological zones and planting systems?
4. What are the implications of these improvements for large-scale monitoring and sustainability reporting in the cocoa supply chain?

During the internship you will:

• Preprocess and harmonize Sentinel-1, Sentinel-2, and ancillary data for model training
• Explore RAG architectures for geospatial classification (e.g., embedding retrieval, context augmentation)
• Train and evaluate ML models using the ofi cocoa farm polygon database
• Benchmark RAG-enhanced models against traditional machine learning baselines
• Produce spatially explicit cocoa distribution maps for Côte d’Ivoire
• Document model performance, uncertainty, and generalization across regions
• Present results to forecast and sustainability teams of ofi

This internship is ideal for a student who:

• Is an MSc student in the final phase of their studies (e.g. Remote Sensing, Geomatics, Data Science, AI, Environmental Sciences or similar)
• Has a strong interest in geospatial data, machine learning and sustainability
• Has experience with Python and geospatial workflows (e.g. Google Earth Engine, QGIS, GDAL, Rasterio)
• Is familiar with deep learning frameworks such as PyTorch or TensorFlow
• Is curious about advanced ML concepts such as representation learning, transformers or RAG
• Can work independently and manage complex datasets and workflows
• Is comfortable working with confidential information and willing to sign a Non-Disclosure Agreement (NDA)