This dataset supports individual tree-crown object detection from high-resolution geospatial image chips in agroforestry and agricultural settings in India. Each release includes GeoTIFF imagery, clean COCO-format annotations for a single category (tree), image-level metadata and manifests, and multiple train/evaluation split definitions—including a fixed train / in-distribution test partition and a non–distribution-shift protocol for geographic robustness. The split metadata includes state, region, and zone fields to support geographic evaluation beyond random image splits.
This release contains GeoTIFF image chips and COCO-format tree-crown annotations for individual tree detection in agroforestry landscapes.
geotiffs/: 13,492 GeoTIFF image chips.annotations/instances_india_tree_detection_coco_clean.json: clean COCO annotations.metadata/india_geotiff_release_manifest.csv: one row per image with metadata.splits/: train/evaluation split files used by the benchmark and non-distribution-shift baseline.samples/: small reviewer-inspection sample.croissant.json: machine-readable dataset metadata.treeThe release includes:
india_train.* and india_id_test.*: fixed released train and ID-test partition.non_distshift_train_val_test.csv: non-distribution-shift protocol used by benchmark training jobs.traditional_cv_*.txt: conventional train/validation/test aliases.The split metadata includes state, region, and zone fields for geographic evaluation.
This dataset release is licensed under Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0).
Orthophoto and imagery underlying the GeoTIFF chips remain subject to the original acquisition and provider agreements summarized in croissant.json (prov:wasDerivedFrom). The CC BY-NC license applies to this release as packaged; confirm compatibility before commingling with other sources or uses outside non-commercial terms where NC applies.