OAM-TCD: A globally diverse dataset of high-resolution tree cover maps

This repository contains files for the OAM-TCD dataset. This repository contains:

• GeoTIFF format images (images.tar.gz)
• Semantic segmentation annotation masks (masks.tar.gz)
• MS-COCO annotation files (train.tar.gz and test.tar.gz)
• Associated metadata for images in each split

Images are named <oam_id>_<image_id>.tif

For more information, see our arXiv paper here: https://arxiv.org/abs/2407.11743

We recommend that you download the dataset via HuggingFace Hub and we provide a utility to convert the dataset (including folds) to disk in our repository. This archive is provided mainly for long-term availability and reference.

The data are split into three groups depending on image license. The vast majority of the data are CC BY 4.0 licensed (approx. 90%), with smaller portions as CC BY-NC 4.0 and CC BY-SA 4.0. These subsets have the zip extension '-nc' and '-sa' respectively. All CC BY-SA images are in the test set.

Additionally, we provide dataset split indices that can be used for 5-fold cross-validation. To avoid duplication, we do not provide separate annotation files for each fold. You can find these indices in the JSON files in the metadata using the image_id as a key. Each image is given a validation_fold which is an integer in [0,4], a value of -1 indicates that the image belongs to the holdout dataset and should not be used for training with this split arrangement.

All images in the dataset are courtesy of contributors of the Open Imagery Network via Open Aerial Map.