Tea Plantation Mapping in Africa: A Multi-Country Satellite Imagery and Ground-Truth Dataset for Training, Validation, and Testing Deep Learning Model

This dataset supports tea plantation mapping using satellite imagery and machine/deep learning, targeting researchers in geospatial machine learning, remote sensing, and cropland mapping in data-deficient regions, particularly in Africa. It comprises satellite image tiles (.tif) and corresponding vector ground-truth annotations (.gpkg) for tea plantations across eight African countries, split into training/validation data (Kenya, Ethiopia, Rwanda, and South Africa) and out-of-distribution testing data (Burundi, Malawi, Uganda, and Zimbabwe) to evaluate model generalization to unseen regions. The dataset is organized as follows: data/training/ — image and label pairs for model training (Kenya, Ethiopia, Rwanda, South Africa)     data/validation/ — image and label pairs for model validation (Kenya, Ethiopia, Rwanda, South Africa)   data/testing/ — held-out image and label pairs for cross-country generalization testing (Burundi, Malawi, Uganda, Zimbabwe)   script/ — Python script to train, validate, and test a tea plantation segmentation model on the provided data; can also be adapted to train custom models This dataset and code accompany a manuscript currently under review at XXX journal. Files are embargoed and will be made publicly accessible upon publication (expected by the end of August 2026). Please cite both this dataset and the associated paper (citation to be added upon publication) when using these materials. Intended uses include benchmarking cropland/tree-crop classification models, transfer learning and domain generalization studies across African agroecological zones, and developing data-efficient or label-efficient mapping approaches for data-deficient regions.