Overview

The Mapping Africa project has developed models and data geared towards field boundary mapping in various countries in Africa. The largest of these efforts was through the Lacuna Fund-based project led by Farmerline and collaboration with Spatial Collective to develop A Region-Wide, Multi-Year Set of Field Boundary Labels for Africa (the Lacuna labels), which are now hosted on both the Registry of Open Data on AWS and Zenodo.

In addition to that dataset, an additional set of ~5000 labels collected through various Mapping Africa project activities are to be made available. These data were combined with the Lacuna labels to train a modified U-Net model (described in Khallaghi et al, 2025) that has been applied to map several countries, including Zambia, Tanzania, Angola, Ghana, and Nigeria.

The goal of this project is to integrate these datasets and models with the broader Fields of the World project. This integration will revolve around two primary efforts:

1. Integrate the Lacuna+ labels with the existing FTW labels.
2. Train and evaluate models using various combinations the integrated datasets. Models will include the existing Mapping Africa U-Net as well as FTW's U-Net variant, and potentially others.

Set-up

We require ftw-tools to be installed (currently part of ftw-baselines), which requires python 3.10-3.12. Using pyenv to manage and set up the environment:

pyenv install -v 3.12.10
pyenv virtualenv 3.12.10 ftw-mapafrica
pyenv activate ftw-mapafrica
python -m pip install --upgrade pip

And then run pip install -e . to install the package in editable mode.

Datasets

We'll get the Lacuna+ labels from our own HPC storage, and the FTW dataset using the FTW cli.

ftw download 
ftw data download -o ~/data/labels/cropland/

These will be handled by a dataset class and data module based on FTW's as well as modifications from torchgeo.