Cocoa Probability Maps for Nigeria 2020

Cocoa probability map generated as part of the Master’s thesis “Generalization of Cocoa Mapping Across Countries Using Foundation Models and Sparse Labels”, Robotics, Cognition, Intelligence (RCI), Technical University of Munich (TUM).

The main dataset, 2020-nigeria-cocoa-probs.zip, consists of multiple GeoTIFF tiles covering Nigeria. Each tile is stored in its local coordinate reference system (CRS), as defined in the metadata of the respective file. All tiles have a spatial resolution of 10 m.

Raster pixel values represent the model-predicted probability of cocoa presence, encoded as integers:

• 0 → 0% probability

• 250 → 100% probability

• 255 → No data

The values were linearly scaled from the model’s continuous 0–1 probability output. The full methodology, model architecture, training procedure, and data sources are described in the aforementioned Master’s thesis.

Supplementary Files

Three auxiliary prediction sets used in the generation of the final cocoa probability map are also provided for transparency and reproducibility:

• 2020-aef-predictions.zip – Predictions from the AlphaEarth-based model.

• 2020-croma-predictions.zip – Predictions from the CROMA foundation model.

• 2020-baseline-cnn-predictions.zip – Predictions from the baseline CNN model.

These datasets are included as supplementary material to document intermediate outputs but are not required for interpreting or using the final cocoa probability map.