{"success":true,"data":{"author":"Weecology lab, University of Florida","bucketAccessType":"public","bucketName":"weecology/neon-tree-crowns-dta","bucketPrefix":"","category":"Environmental Science","cloudProvider":"huggingface","createdAt":"2026-11-16T02:38:32.527Z","description":"A unified, species-labelled tree-crown polygon set covering 38 National Ecological Observatory Network (NEON) field sites, from the Weecology lab. It has 41,738 crowns for 39,702 individual trees across 234 NEON taxonIDs. About 81% are algorithmic: the DeepTreeAttention (DTA) pipeline ran DeepForest crown detection on NEON airborne RGB imagery and matched each detection to a NEON Vegetation Structure (VST) stem record for its species, falling back to a small square around the surveyed stem when nothing matched. The other 19% were drawn by hand (5,076 bounding boxes and 2,870 polygons, the polygons often repeated over up to 5 years), and the authors recommend training on algorithmic crowns and evaluating on hand-annotated ones. Columns record crown_source, detection_method, data_source (ForestGEO, NEON, IFAS, megaplot and others), native UTM zone and year. One GeoPackage, neon_crowns_dta.gpkg, about 30 MB, in EPSG:4326; read with geopandas or the neoncrops helper. Released CC BY 4.0; the underlying NEON imagery and field data are also CC BY 4.0.","displayTitle":"NEON Tree Crowns, DeepTreeAttention Edition: 41,738 Species-Labelled Crowns at 38 NEON Sites","id":"820901da-2e0e-5ffb-9bba-b9577fbedf37","license":"CC-BY-4.0","licenseEvidenceUrl":"https://huggingface.co/datasets/weecology/neon-tree-crowns-dta","objectCountEstimate":3,"probeStatus":"ok","region":"us","schemaCard":{"accessHints":[{"language":"python-datasets","snippet":"from datasets import load_dataset\nds = load_dataset(\"weecology/neon-tree-crowns-dta\")"},{"language":"python-hub","snippet":"from huggingface_hub import snapshot_download\nsnapshot_download(repo_id=\"weecology/neon-tree-crowns-dta\", repo_type=\"dataset\")"}],"caveats":["HuggingFace metadata: author=weecology, size_categories:10K<n<100K, downloads=102"],"format":"mixed","generatedAt":"2026-11-16T02:38:38.097Z","layout":{"example":"neon_crowns_dta.gpkg","pattern":"{repoId}/{filename}"},"mountPath":"/mnt/archive","sampleHeaders":[{"bytes":"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","encoding":"hex","key":"neon_crowns_dta.gpkg","truncatedAt":512}],"sampleObjects":["neon_crowns_dta.gpkg"],"sidecars":[{"bytes":"---\npretty_name: NEON Tree Crowns (DTA + hand-annotated)\ntags:\n  - neon\n  - tree-crowns\n  - remote-sensing\n  - aerial\n  - geospatial\n  - ecology\n  - biodiversity\nsize_categories:\n  - 10K<n<100K\n---\n\n# NEON Tree Crowns — DTA edition\n\nA unified, species-labeled tree-crown polygon set covering 38 NEON sites,\ncombining algorithmic crowns from the DeepTreeAttention (DTA) pipeline with\nhand-annotated bounding boxes and polygons curated by the Weecology lab.\n\n| | |\n|---|---|\n| **rows** | 41,738 crowns |\n| **individuals** | 39,702 unique trees |\n| **species** | 234 (NEON `taxonID`) |\n| **sites** | 38 NEON sites |\n| **format** | single GeoPackage (`neon_crowns_dta.gpkg`) |\n| **CRS** | EPSG:4326 (WGS84). Native UTM zone in `crs_epsg`. |\n\n## Quickstart\n\n```python\nfrom neoncrops.fetch_crowns import fetch_crowns\ngdf = fetch_crowns()  # downloads + caches the gpkg, returns a GeoDataFrame\n```\n\n## Where the crowns come from\n\nCrowns in this dataset come from two complementary sources, recorded in the\n`crown_source` column. Most of the dataset (about 81%) is algorithmic: the\nDTA pipeline ran DeepForest tree detection on NEON RGB imagery and matched\neach detection to a NEON Vegetation Structure (VST) stem record to attach a\nspecies label. When no detection matched a known stem, the pipeline falls\nback to a small square around the stem's surveyed location so the tree is\nnot lost. The remaining 19% are crowns that humans drew by hand on the\nsame imagery: some as axis-aligned bounding boxes (`hand_bbox`), and a\nsmaller set as careful polygon outlines repeated across multiple years\n(`hand_polygon`).\n\n| `crown_source` | rows | what it is |\n|---|---:|---|\n| `algorithmic` | 33,792 | DeepForest detections (`detection_method='deepforest'`) or point-buffered fallbacks (`detection_method='fallback'`), matched to a NEON VST stem |\n| `hand_bbox` | 5,076 | manually drawn axis-aligned bounding boxes |\n| `hand_polygon` | 2,870 | manually drawn polygon outlines, often the same tree across multiple years |\n\nThe `data_source` column records the originating sub-dataset within each\ngroup (ForestGEO, neon, IFAS, clark, megaplot, arboretum, townsend, fricker,\nhand_annotated).\n\n## Hand-annotated crowns: the recommended evaluation set\n\nThe 7,946 hand-annotated rows (`crown_source IN ('hand_bbox', 'hand_polygon')`)\nare the highest-quality crowns in the dataset. A human looked at every one of\nthem on the imagery, so they are not subject to the detection mistakes,\nmismatches, or buffer approximations that the algorithmic crowns can have.\nMost users should treat them as the held-out evaluation set:\n\n- **Train on `algorithmic`**, evaluate on `hand_*`.\n- The `hand_polygon` subset is multi-temporal (mean ~3.3 years per individual,\n  up to 5 years), which makes it useful for testing temporal generalization.\n- Reserve `hand_bbox` if your task only needs bounding boxes, or use both.\n\n```python\ngdf = fetch_crowns()\ntrain = gdf[gdf[\"crown_source\"] == \"algorithmic\"]\neval_set = gdf[gdf[\"crown_source\"].isin([\"hand_bbox\", \"hand_polygon\"])]\n```\n\n## Schema\n\n| column | dtype | notes |\n|---|---|---|\n| `crown_id` | str | row-unique. `{individual}` for algorithmic rows, `{individual}_{anno_year}` for hand rows |\n| `individual` | str | NEON individual ID (e.g. `NEON.PLA.D02.SCBI.00011`) |\n| `siteID` | str | NEON 4-letter site code |\n| `plotID` | str | NEON plot ID (nullable) |\n| `taxonID` | str | NEON `taxonID` after harmonization |\n| `taxonID_original` | str | pre-harmonization `taxonID` |\n| `scientificName` | str | only populated for hand crowns |\n| `crown_source` | str | `algorithmic` \\| `hand_bbox` \\| `hand_polygon` |\n| `data_source` | str | originating dataset name |\n| `detection_method` | str | `deepforest` \\| `fallback` \\| `hand_annotated` |\n| `anno_year` | int | year of imagery used for hand annotation (hand rows only) |\n| `detection_year` | int | year of imagery used by the detector (algorithmic rows only; missing for IFAS) |\n| `height`, `CHM_height` | float | field-measured stem height; CHM-derived crown height (algorithmic only) |\n| `stemDiameter`, `canopyPosition`, `plantS","encoding":"utf8","key":"README.md","name":"README.md","size":6644,"truncated":true}],"sourceObjectKey":"neon_crowns_dta.gpkg","starterCell":"from huggingface_hub import hf_hub_download\npath = hf_hub_download(repo_id=\"weecology/neon-tree-crowns-dta\", filename=\"neon_crowns_dta.gpkg\", repo_type=\"dataset\")\nprint(path)","starterSamplePath":"neon_crowns_dta.gpkg","totalObjects":3},"summary":"A unified, species-labelled tree-crown polygon set covering 38 National Ecological Observatory Network (NEON) field sites, from the Weecology lab.","title":"NEON Tree Crowns, DeepTreeAttention Edition: 41,738 Species-Labelled Crowns at 38 NEON Sites","updatedAt":"2026-11-16T17:11:25.677Z","visibility":"public","websiteUrl":"https://huggingface.co/datasets/weecology/neon-tree-crowns-dta"}}