Geoalert/neon-polygons

NEON tree crown polygons

Polygons of individual tree crowns on the National Ecological Observatory Network's 10 cm RGB orthomosaics, 5,144,094 of them, made by prompting SAM 3 with crown boxes that already exist: the hand-drawn boxes of NeonTreeEvaluation and the predicted boxes of DeepForest's NEON Tree Crowns. They are pseudo-labels: no human has checked them yet (see Versions).

Each image comes whole, with all its crowns, as in GIS: the image as a Cloud-Optimized GeoTIFF, its crowns as one GeoParquet file in the image's CRS (UTM), and a preview with the crowns drawn for this page's viewer.

Subset (config)	Split	Images	Sites	Polygons	Image size (px, max)	"No label here" means	Size
val-hand	validation	194	22	6,633	400	background (exhaustive boxes)	0.1 GB
train-hand	train	20	14	23,465	10,000	unknown (may hold a tree)	0.6 GB
deepforest-5pct	train	245	36	5,113,996	10,000	unknown (may hold a tree)	14.9 GB

Coming: deepforest-95pct, the other ~95% of the NEON Tree Crowns' boxes (about 99.6M), labels only, in parts.

What to use it for

Use	Data	Note
train or fine-tune a crown segmentation model	train-hand, deepforest-5pct	pseudo-labels: filter by score and the flags; chips are cut by a recipe (below)
evaluate crown segmentation	val-hand, once reviewed by a human	until then it is SAM 3's output: scoring SAM-like models against it is circular
ecology and GIS: crown area, size distributions, joins with NEON's other products	labels/ of any subset	native UTM, one file per 1 km tile or chip
train or evaluate crown detection (boxes)	source_box_* of val-hand and train-hand	the hand-drawn boxes, as they are

For crown detection, the source boxes of val-hand and train-hand (source_box_*) are NeonTreeEvaluation's hand-drawn boxes, the only human labels here: use them rather than the polygons' bounding boxes, val-hand's as the exhaustive test set.

How they were made

Stage 1 of AIST, an agentic instance-segmentation toolkit (the code will follow):

1. each box is grown by 1.3 about its centre (real boxes cut crowns short, by the validation) and given to SAM 3's tracker with a click at its centre;
2. at two image scales (1 and 2); the two masks are averaged (the mean of their signed distance maps);
3. the result is simplified to 1 px. Where SAM 3 returns no mask (2 of 5.1M boxes), the box is kept (form box).

SAM 3 ran as TensorRT engines on one RTX 3090 (deepforest-5pct: 5.1M boxes in 10.1 h). The runs' configs, engine ids and timings are in provenance/.

How good are they

Against hand-drawn crown polygons (the method's validation, before this release), the polygons fit the crowns better than the boxes they come from:

Boxes	Trees with a hand polygon	Mean IoU, the box itself	Mean IoU, the polygon
DeepForest's predicted boxes	576 (5 sites)	0.633	0.686
NeonTreeEvaluation's hand boxes	67	0.640	0.720

Known failure modes:

• Overlaps in dense stands. The prompt box is grown by 1.3 (real boxes cut crowns short), so neighbours can claim the same canopy: overlap > 0.3 for 2.8% of val-hand's polygons, most at SCBI (8.2%), SERC (7.4%) and TALL (6.7%). Filter with overlap, or resolve the overlaps after download.
• Reaching outside the box (leak > 0.3): 3.2% of val-hand, most at NIWO (8.9%) and ONAQ (9.4%), often clumps.
• The boxes' own errors stay: a polygon exists for every box and only for boxes. DeepForest's boxes have a precision of about 0.73 and a recall of about 0.79 on NeonTreeEvaluation; trees it missed have no polygon, and its false boxes get a polygon too (deepforest-* "no label here" means nothing).
• NEON's 2021+ mosaics smear crowns in places; the DeepForest tiles here are from 2018–2019.

Every polygon carries score (SAM 3's predicted IoU, the best single flag with real boxes) and the GT-free flags, so the doubtful ones can be dropped or reviewed.

Layout

README.md
index/<subset>.parquet                        the configs above: one row per image, previews embedded
<subset>/dataset.json                         the subset as an AIST dataset (images, sizes, CRS, transforms)
<subset>/images/<SITE>/<image>.tif            the whole image: a Cloud-Optimized GeoTIFF with overviews
<subset>/labels/<SITE>/<image>.parquet        all its crowns: GeoParquet 1.1, the image's CRS
provenance/<run>/run.json, summary.json, export.json

• Images. val-hand and train-hand keep their source's class: lossless sources as lossless COGs (deflate), JPEG sources as JPEG at quality 95. deepforest-5pct is JPEG at quality 90 throughout (NEON's tiles are mostly JPEG already). Internal overviews and 512 px tiles: any zoom level streams by HTTP range requests.

• The index (what this page's viewer shows) has preview and preview_overlay (the whole image at 1024 px, plain and with the crowns drawn), thumbnail and thumbnail_overlay (the same at 256 px). The overlays colour each crown by its score: red at 0.6 and below, yellow at 0.78, green at 0.92 and above, grey without a score; a crown too small to outline at that size is filled only. Per image: image_id, site, year, month, width, height, gsd_m, crs_epsg, transform (rasterio's a, b, c, d, e, f), bounds, centroid_lon, centroid_lat, the counts (n_labels, n_polygons, n_auto, n_none), score_median, the flags' shares (share_leak, share_overlap: > 0.3; share_multipart), label_version, the source (source_product, source_release, source_file, source_url, source_size, source_crc32c, source_sha256) and the paths (image_path, labels_path, with their sizes). The paths are plain strings: load_dataset does not download the images.

• The labels, one row per crown:

  Column	What
  image_id, crown_id	the image; the crown's stable id, <image_id>/<source box id>
  geometry, bbox, area_m2	the polygon (WKB, the image's UTM zone, mm precision), its bbox (the GeoParquet covering), its area
  form, label, status	polygon, or box where SAM 3 returned no mask (status none: the box is kept); Tree; auto
  score, score_scale1, score_scale2, object_score	SAM 3's predicted IoU: the polygon's (the mean of the two scales'), each image scale's, and the object score
  source_box_xmin, source_box_ymin, source_box_xmax, source_box_ymax	the source box, the crown's prompt (provenance, not a second label), in the same CRS
  source_box_origin, source_box_score, source_box_height_m	where the box comes from (hand-image or deepforest); DeepForest's score and its tree height (from NEON's LiDAR)
  box_iou, fill, leak, overlap, parts, compactness	the GT-free flags: the polygon's bbox IoU with its source box; its area over the box's; its share outside the box; the largest share of it another polygon claims; its parts; 4πA/P²
  origin, method, model, label_version	sam3-tracker, M0ms_bscale1.3, the engine's id; the version in which the row last changed

  Three train-hand tiles have no georeference in their source files (2019_DSNY_5_452000_3113000_image_crop, 2019_YELL_2_528000_4978000_image_crop2, 2019_YELL_2_541000_4977000_image_crop): their crowns are in image pixels (y down), with a null CRS, and area_m2 and the index's georeference columns are null.

Usage

The index (previews and attributes, no large downloads):

from datasets import load_dataset

index = load_dataset("Geoalert/neon-polygons", "deepforest-5pct", split="train", revision="v1.0.0")

One image's crowns, with GeoPandas:

import geopandas as gpd
from huggingface_hub import hf_hub_download

path = hf_hub_download("Geoalert/neon-polygons", "deepforest-5pct/labels/ABBY/ABBY_2019_551000_5071000.parquet",
                       repo_type="dataset", revision="v1.0.0")
crowns = gpd.read_parquet(path)  # EPSG:32610
good = crowns[(crowns.score > 0.8) & (crowns.overlap < 0.3)]

All the crowns of a subset, with DuckDB (no download of the images):

SELECT split_part(image_id, '_', 1) AS site, count(*) AS crowns, median(area_m2) AS median_area_m2
FROM 'hf://datasets/Geoalert/neon-polygons@v1.0.0/deepforest-5pct/labels/**/*.parquet'
GROUP BY site ORDER BY crowns DESC;

A whole subset, images included (dataset.json describes it for AIST, whose code will follow):

from huggingface_hub import snapshot_download

root = snapshot_download("Geoalert/neon-polygons", repo_type="dataset", revision="v1.0.0", allow_patterns=["train-hand/*"])

Training chips

No chips are shipped: the images are whole. To cut chips (e.g. 1000 x 1000 px, which divides the 10,000 px tiles and cuts about 9% of the crowns), read a window of the COG and clip the crowns to it:

import geopandas as gpd, rasterio
from rasterio.windows import Window, bounds
from shapely import affinity, box

with rasterio.open(image_path) as r:  # the image's COG
    win = Window(0, 0, 1000, 1000)
    rgb = r.read(window=win)  # (3, 1000, 1000)
    crowns = gpd.read_parquet(labels_path)
    clip = crowns.clip(box(*bounds(win, r.transform)))
    t = ~r.window_transform(win)  # CRS -> chip pixels
    pixel_polygons = [affinity.affine_transform(g, [t.a, t.b, t.d, t.e, t.c, t.f]) for g in clip.geometry]

Versions

Tags vMAJOR.MINOR.PATCH: the major is the stage, the minor the coverage, the patch fixes; main is the latest. Pin a tag for reproducible work.

• v1.0.0 stage 1: val-hand, train-hand, deepforest-5pct, with images and previews
• v1.1.0 … v1.N.0 deepforest-95pct, stage 1, labels only, in parts (a few sites per part)
• v1.N.0 stage 1 complete: all of the NEON Tree Crowns' 104.7M boxes
• v1.x val-hand reviewed by a human: the first NEON crown polygon test set
• v2.0.0 stage 2 for val-hand and train-hand: clumps split, missed trees added, false boxes removed
• v2.x stage 2 for the DeepForest subsets; the final version: corrected polygons (the method is open)

Changelog

• 2026-11-05: v1.0.0: the first release, stage 1: val-hand, train-hand, deepforest-5pct.

License and attribution

CC BY 4.0. Attribute this dataset and its sources:

• NEON. The imagery is NEON's, released under CC BY 4.0. "The National Ecological Observatory Network is a program sponsored by the U.S. National Science Foundation and operated under cooperative agreement by Battelle." "This material is based in part upon work supported by the National Ecological Observatory Network (NEON), a program sponsored by the U.S. National Science Foundation (NSF) and operated under cooperative agreement by Battelle." No endorsement by NEON is implied.
  • deepforest-5pct: NEON (National Ecological Observatory Network). High-resolution orthorectified camera imagery mosaic (DP3.30010.001), RELEASE-2026. https://doi.org/10.48443/8vq2-s021. Dataset accessed from https://data.neonscience.org/data-products/DP3.30010.001/RELEASE-2026 on November 1, 2026.
  • val-hand, train-hand: NEON's RGB tiles as NeonTreeEvaluation distributes them (below).
• NeonTreeEvaluation (the hand boxes; the chips and training tiles): Weinstein B., Marconi S., White E. Data for the NeonTreeEvaluation Benchmark (0.2.2), Zenodo, 2022, https://doi.org/10.5281/zenodo.5914554, CC BY 4.0; and Weinstein B. G. et al. (2021) A benchmark dataset for canopy crown detection and delineation in co-registered airborne RGB, LiDAR and hyperspectral imagery from the National Ecological Observation Network. PLoS Comput Biol 17(7): e1009180, https://doi.org/10.1371/journal.pcbi.1009180.
• NEON Tree Crowns (DeepForest's boxes): Weinstein B. et al. NEON Tree Crowns Dataset (0.0.1), Zenodo, 2020, https://doi.org/10.5281/zenodo.3765872, CC BY 4.0; and Weinstein B. G. et al. (2021) A remote sensing derived data set of 100 million individual tree crowns for the National Ecological Observatory Network. eLife 10: e62922, https://doi.org/10.7554/eLife.62922.
• SAM 3. The polygons were made with SAM Materials (SAM License, 2025-11-19): Carion N. et al. (2025) SAM 3: Segment Anything with Concepts, https://arxiv.org/abs/2511.16719. This dataset contains no SAM 3 weights or code.

Changes made: the boxes were converted to polygons by SAM 3 (above); the images were re-encoded as COGs (deepforest-5pct: JPEG at quality 90) and previews added.

Authors

Mikhail Kotyushev, Geoalert. Questions, problems and fixes: the Community tab.

Citation

@misc{neon_polygons_2026,
  title        = {NEON tree crown polygons},
  author       = {Mikhail Kotyushev},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/Geoalert/neon-polygons}},
  note         = {Version 1.0.0}
}

Please also cite NEON, NeonTreeEvaluation, the NEON Tree Crowns and SAM 3 (above).