{"success":true,"data":{"author":"Tseng, Zvonkov, Nakalembe, Kerner (NASA Harvest, University of Maryland); TACO repackaging by ISP, Universitat de Valencia","bucketAccessType":"public","bucketName":"isp-uv-es/cropharvest-taco","bucketPrefix":"","category":"Agriculture & Environment","cloudProvider":"huggingface","createdAt":"2026-11-16T14:11:25.737Z","description":"CropHarvest (NeurIPS 2021 Datasets and Benchmarks) is a global crop-type and cropland mapping dataset that pools 70,213 labelled agricultural reference points from 25 source datasets on six continents. Every point carries a 12-month time series of 18 variables at its 10 m pixel: Sentinel-1 VV and VH, eleven Sentinel-2 surface reflectance bands, ERA5 temperature and precipitation, SRTM elevation and slope, and NDVI. All points have a binary crop or non-crop label and 29,749 also a ten-class crop category, so the benchmark can measure transfer between countries and continents. Each sample is one pixel (rasters are 1x1), so it suits per-pixel temporal models rather than CNNs; this copy groups variables into seven slots by unit and calibration with twelve monthly frames each. Splits are 58,998 train, 10,802 validation and 413 test. About 3.5 GB in TACO format, read with tacoreader, or use the cropharvest Python package. Released by NASA Harvest under CC BY-SA 4.0; cite Tseng et al. 2021.","displayTitle":"CropHarvest: Global Crop / Non-Crop Point Time Series from 25 Datasets (TACO)","id":"86905600-d771-54a4-b7e6-8e4b889f48ab","license":"CC-BY-SA-4.0","licenseEvidenceUrl":"https://github.com/nasaharvest/cropharvest","objectCountEstimate":13,"probeStatus":"ok","referenceUrl":"https://openreview.net/forum?id=JtjzUXPEaCu","region":"us","schemaCard":{"accessHints":[{"language":"python-datasets","snippet":"from datasets import load_dataset\nds = load_dataset(\"isp-uv-es/cropharvest-taco\")"},{"language":"python-hub","snippet":"from huggingface_hub import snapshot_download\nsnapshot_download(repo_id=\"isp-uv-es/cropharvest-taco\", repo_type=\"dataset\")"}],"caveats":["HuggingFace metadata: author=isp-uv-es, license:cc-by-sa-4.0, size_categories:10K<n<100K, downloads=113"],"format":"mixed","generatedAt":"2026-11-16T14:11:34.822Z","layout":{"example":"assets/logos.png","pattern":"{repoId}/{filename}"},"mountPath":"/mnt/archive","sampleHeaders":[{"bytes":"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","encoding":"hex","key":"assets/logos.png","truncatedAt":512}],"sampleObjects":["assets/logos.png","assets/sample_0.png","assets/sample_1.png","assets/sample_2.png","assets/sample_3.png"],"sidecars":[{"bytes":"---\nlicense: cc-by-sa-4.0\npretty_name: \"CropHarvest\"\nsize_categories:\n- 10K<n<100K\ntask_categories:\n- image-classification\ntags:\n- earth-observation\n- remote-sensing\n- taco\n---\n\n<p align=\"center\"><img src=\"assets/logos.png\" width=\"546\" alt=\"Image and Signal Processing Group, Universitat de València · ELLIOT\"></p>\n\n# CropHarvest\n\n> **This is a repackaging, not a new dataset.** It is **CropHarvest** by NASA Harvest / University of Maryland (Tseng et al.), ESA Copernicus / Sentinel-1 and Sentinel-2, converted to TACO. Pixel values and labels are kept as released except where the description below says otherwise. All credit belongs to the original authors: if you use it, please cite them and follow their licence.\n>\n> [original dataset](https://github.com/nasaharvest/cropharvest) · [paper](https://openreview.net/forum?id=JtjzUXPEaCu) · licence: CC-BY-SA-4.0\n>\n> Repackaged into TACO by the [Image and Signal Processing Group (ISP)](https://isp.uv.es/), Universitat de València, within the [ELLIOT](https://elliot-ai.eu) project.\n\n## Citation\n\nPlease cite the original work:\n\n```bibtex\n@inproceedings{tseng2021cropharvest,\n  title     = {CropHarvest: A global dataset for crop-type classification},\n  author    = {Tseng, Gabriel and Zvonkov, Ivan and Nakalembe, Catherine Lilian and Kerner, Hannah},\n  booktitle = {NeurIPS Datasets and Benchmarks Track},\n  year      = {2021}\n}\n```\n\n## About the data\n\n70213 agricultural reference points pooled from 25 source datasets across six continents, each carrying a 12-month time series of 18 per-pixel variables: Sentinel-1 VV/VH, eleven Sentinel-2 surface reflectance bands, ERA5 temperature and precipitation, SRTM elevation and slope, and NDVI. Every point has a binary crop/non-crop label, and 29749 of them also a ten-class crop category. It is a global crop dataset rather than a single-country one, so cross-continent transfer can be measured on it.\n\n**70,213 samples** · splits: `test` 413 · `train` 58,998 · `validation` 10,802 · tasks: `classification`\n\nPackaged as [TACO v3](https://asterisk.coop/taco/spec/).\n\n<details>\n<summary>Full description</summary>\n\n**Point data, not imagery.** Each sample is one 10 m pixel and every raster is 1x1 px.\n- Spatial crops, resizes, flips and tiling windows are meaningless here, and a convolutional backbone has nothing to convolve over. What it supports is a per-pixel model over time, 12 steps.\n- Filter on the `point-data` keyword when mixing it with image datasets.\n- There are 62 files per sample, about 4.35 million archive entries in total, so per-sample IO is dominated by file count rather than bytes.\n\n**Layout.** The 18 variables span five units and two calibrations, and are written as seven slots grouped by (calibration, scale, unit) rather than one raster: Sentinel-2 is reflectance x 10,000 while NDVI is a plain ratio, the same unit under a different calibration, so even those two cannot share a slot. The five monthly groups are twelve-frame raster series; folding the months into the band axis would give Sentinel-2 132 bands and invite it being read as hyperspectral. Elevation and slope are constant across the year and are stored once, verified per sample.\n\n**Labels.** They join to the arrays on (index, dataset), exactly for all 70213. labels.geojson holds a further 24,973 points the release exported no features for, which are not samples here.\n\n**Splits.** The official is_test split, preserved unaltered: 69800 train against 413 test. It is lopsided because the release's test set is a small curated evaluation set rather than a random holdout. The release ships no val part, so one is carved out of train on whole spatial blocks -- 0.1-degree cells merged into connected components, 25,934 of them -- never out of test.\n\n</details>\n\n## Getting started\n\n```bash\ngit clone --recursive https://github.com/OscarPellicer/taco\npip install -e \"taco/python[ml]\"     # builds the reader: C++23, CMake, Ninja, pkg-config, libcurl >= 7.83, OpenSSL >= 3\n```\n\nRead it straight from the Hub:\n\n```python\nfrom huggingface_hub import snapshot_download\nfrom taco.ml import Dat","encoding":"utf8","key":"README.md","name":"README.md","size":6500,"truncated":true}],"sourceObjectKey":"assets/logos.png","starterCell":"from huggingface_hub import hf_hub_download\npath = hf_hub_download(repo_id=\"isp-uv-es/cropharvest-taco\", filename=\"assets/logos.png\", repo_type=\"dataset\")\nprint(path)","starterSamplePath":"assets/logos.png","totalObjects":13},"summary":"CropHarvest (NeurIPS 2021 Datasets and Benchmarks) is a global crop-type and cropland mapping dataset that pools 70,213 labelled agricultural reference points from 25 source…","title":"CropHarvest: Global Crop / Non-Crop Point Time Series from 25 Datasets (TACO)","updatedAt":"2026-11-16T16:13:58.518Z","visibility":"public","websiteUrl":"https://huggingface.co/datasets/isp-uv-es/cropharvest-taco"}}