minute1028/forestllava-dataset

Forest-LLaVA Multimodal Tree-Species Dataset

Forest-LLaVA is a multimodal remote-sensing dataset for tree-species recognition and structured vision-language research. Each record is indexed by a numeric sample_id from the US subset of GlobalGeoTree and is linked to a 60 m × 60 m patch from NAIP Optical, Sentinel-2 MSI and Sentinel-1 SAR data, four-level taxonomic labels and geographic/environmental records.

The repository contains the complete image archives for the paper-aligned US-50 and US-200 configurations, their split CSV files, and the corresponding labels and Geo records. Large files are stored with Git LFS/Xet.

Dataset configurations

Configuration	Train	Validation	Test	Species	Purpose
US-50 formal	36,312	4,552	4,551	50	Main class-stratified benchmark
US-50 spatial hold-out	35,611	4,806	4,686	50	0.25° blocks with a 1 km train–evaluation buffer
US-200 common-OMS	138,721	17,248	17,386	200	Expanded-class benchmark with common Optical–MSI–SAR coverage

The spatial hold-out split reuses the US-50 image pool; it is defined by its own CSV files and does not require a second copy of the images. US-200 includes the US-50 records, so some image patches are intentionally duplicated between the two configuration directories to support independent downloads.

Repository layout

splits/
├── us50_formal/{train,val,test}.csv
├── us50_spatial_holdout/{train,val,test}.csv
└── us200_common_oms/{train,val,test}.csv

annotations/
├── labels_json.tar.zst       # {sample_id}.text.json
├── geo_json.tar.zst          # {sample_id}.auxiliary.json
└── SHA256SUMS

images/
├── us50/{optical,multispectral,sar}/
└── us200/{optical,multispectral,sar}/

Each modality directory contains one or more part-*.tar archives and a shard_manifest.csv with the file count, sample-ID range, archive size and SHA-256 checksum. The archives contain TIFF patches at their top level. The numeric prefix before the first underscore in an image filename is the join key for the split CSV and JSON records.

Modalities and archive sizes

Configuration	Optical	MSI	SAR	Total
US-50	1.58 GiB	0.39 GiB	0.17 GiB	2.14 GiB
US-200	6.00 GiB	1.50 GiB	0.66 GiB	8.16 GiB

The US-50 archive contains 45,415 TIFF patches per modality and the US-200 archive contains 173,355 TIFF patches per modality. Because US-200 includes the US-50 records, the two independently downloadable configurations share some underlying samples by design.

Optical patches are derived from USDA Farm Service Agency NAIP aerial imagery. MSI patches are derived from Sentinel-2 Level-2A products, and SAR patches from Sentinel-1 GRD products. The archives contain processed 60 m × 60 m patches rather than the original provider products.

Preview

The following figures show one aligned three-modality patch and the spatial distribution of valid US-50 records. Display enhancements in the patch preview are for visualization only and do not change model inputs.

Download and use

Clone the metadata repository with Git LFS/Xet enabled:

git xet install
git clone git@hf.co:datasets/minute1028/forestllava-dataset

To download only a desired configuration or modality, use the Hugging Face file browser or hf download to select paths under images/us50/ or images/us200/. Extract a shard with a zstd-capable tar implementation and verify it against the corresponding shard_manifest.csv. The split CSVs can be read directly with pandas; sample_id joins images, labels and Geo records.

The two annotation archives cover the union of the US-50 formal, US-50 spatial hold-out and US-200 common-OMS samples (173,355 unique IDs). They are shared by all three split protocols.

Provenance, licence and limitations

Sample identifiers, taxonomic labels and associated geographic records are derived from GlobalGeoTree. GlobalGeoTree is distributed under the Creative Commons Attribution 4.0 International licence; cite the original dataset and paper when using these records.

Optical patches were derived from USDA Farm Service Agency NAIP imagery. MSI and SAR patches contain modified Copernicus Sentinel data processed by the Forest-LLaVA preprocessing pipeline. Please retain the upstream attribution and comply with the terms of the particular product and acquisition year. See licenses/THIRD_PARTY_NOTICES.md for the detailed notices. Those upstream terms control third-party source material; this repository does not grant broader rights.

The observations originate from open records and inherit their geographic sampling pattern. The formal split prioritizes class-balanced comparison and is not spatially independent. The spatial hold-out reduces local proximity between training and evaluation samples and is provided as a complementary spatial-sensitivity protocol, not as a strict cross-region or cross-ecoregion generalization benchmark.

Citation

Please cite the Forest-LLaVA paper, GlobalGeoTree, and the relevant NAIP, Sentinel-1 and Sentinel-2 products. The repository commit and per-shard manifests provide the release version and integrity references for this copy.

中文说明见 README.zh-CN.md。