Note
This is a Hugging Face dataset. For large datasets, ensure huggingface_hub>=1.1.3 to avoid rate limits. Learn more in the Hugging Face integration docs.
Dataset Card for VineLiDAR#
This is a FiftyOne dataset with 10 samples.
Installation#
If you havenât already, install FiftyOne:
pip install -U fiftyone
Usage#
import fiftyone as fo
from huggingface_hub import snapshot_download
# Download the dataset snapshot to the current working directory
snapshot_download(
repo_id="Voxel51/VineLiDAR",
local_dir=".",
repo_type="dataset",
)
# Load dataset from current directory using FiftyOne's native format
dataset = fo.Dataset.from_dir(
dataset_dir=".", # Current directory contains the dataset files
dataset_type=fo.types.FiftyOneDataset, # Specify FiftyOne dataset format
name="VineLiDAR", # Assign a name to the dataset for identification
)
# Launch the App
session = fo.launch_app(dataset)
Dataset Details#
Dataset Description#
VineLiDAR is a collection of ten high-density, RGB-colored 3D LiDAR point clouds acquired by UAV over two commercial vineyard blocks (Vitis vinifera cv. Loureiro, blocks B7 and B9) in Tomiño, Pontevedra, Galicia, Spain. Flights were flown at 20, 30, and 50 meters above ground level (AGL) over two years (2021-2022), across three data-capture phases timed to the vinesâ vegetative growth and harvest stages (September 16, 2021; July 14, 2022; September 8, 2022), using a DJI M300 multi-rotor platform equipped with a DJI Zenmuse L1 LiDAR sensor. The data provides a spatiotemporal record of vineyard canopy and terrain structure, intended as a public benchmark for Precision Agriculture / Precision Viticulture research on woody crops, a domain where public UAV LiDAR datasets are otherwise scarce.
Curated by: Sergio VĂ©lez, Mar Ariza-SentĂs, JoĂŁo Valente (Information Technology Group, Wageningen University & Research)
Funded by: European Commission, H2020 program, FLEXIGROBOTS project (contract number 101017111)
Shared by: Sergio VĂ©lez, Mar Ariza-SentĂs, JoĂŁo Valente, via Zenodo
Language(s): en (documentation/metadata); the data itself is geometric/spectral LiDAR point clouds, not text
License: CC BY 4.0
Dataset Sources#
Repository: Zenodo record 8113105
Paper: VĂ©lez, S., Ariza-SentĂs, M., & Valente, J. (2023). VineLiDAR: High-resolution UAV-LiDAR vineyard dataset acquired over two years in northern Spain. Data in Brief, 109686. https://doi.org/10.1016/j.dib.2023.109686
Uses#
Direct Use#
3D reconstruction and analysis of vineyard canopy and terrain morphology across altitudes and growth stages.
Generating high-precision digital elevation models (DEMs), digital surface/terrain models (DSM/DTM), and canopy height models (CHM) from raw point clouds.
Serving as âground truthâ to validate satellite-derived vineyard models (e.g., Sentinel-2 time series) or to build digital twins.
Providing detailed terrain/canopy geometry for developing and testing UAV/UGV flight paths and navigation algorithms in agricultural robotics.
Studying vineyard development over time by comparing point clouds from the three capture phases (2021 vs. 2022).
Out-of-Scope Use#
The dataset contains no object detections, semantic/instance segmentation, or other annotations (e.g., individual vine, trunk, or grape-bunch labels), so it is not directly usable for supervised detection/segmentation training without additional labeling.
Not intended for real-time onboard navigation without further downstream processing (the released point clouds are post-flight, RTK-corrected products, not live sensor streams).
RGB values were captured during afternoon flights and may show over-exposure or shadowing effects from sunlight; not suitable as a source of calibrated radiometric/reflectance measurements.
Dataset Structure#
This is a flat (non-grouped) FiftyOne point-cloud dataset: media_type="point-cloud", 10 samples, one sample per UAV flight. There are no label fields (no detections, segmentations, or classifications) â each sample is a raw, unlabeled geometric point cloud enriched with flight/acquisition metadata.
Media. Each sampleâs filepath points to a voxel-downsampled .pcd file (~5M points, pcd_view/<name>.pcd) rather than the original full-resolution point cloud, so that flights load quickly and remain responsive in the FiftyOne Appâs interactive 3D viewer. Voxel size was chosen per-flight (voxel_size_m) so that downsampling targets roughly 5,000,000 points regardless of the flightâs footprint or original density; num_points_view records the actual point count after downsampling, and num_points_full records the original point count in the source .laz file for reference. The original full-resolution .pcd files (up to ~1.1 GB, tens of millions of points each) and the source .laz files are kept alongside the parsed dataset but are not referenced by any sample field.
Coordinate handling. The source .laz files use a mix of coordinate systems: the two 2021 flights (epsg_code is null) are stored in a local/relative frame, while the eight 2022 flights are in absolute UTM coordinates (EPSG:32629). To keep point coordinates small and float32-safe for 3D rendering while remaining internally consistent per flight, every point cloud was re-centered on its own bounding-box minimum before being written to .pcd. That per-flight offset (the original bounding-box minimum, in the source CRS) is preserved in the offset_x/offset_y/offset_z fields so the point cloud can be mapped back to its original coordinate system if needed. bounds_min_*/bounds_max_* record the pre-centering bounding box (i.e. bounds_min_* == offset_* by construction), and extent_x/extent_y/extent_z are the derived footprint/height dimensions of each flightâs point cloud. Bounding-box and offset coordinates are stored as flat per-axis scalar fields rather than list fields, consistent with FiftyOne 3D best practices for geometric attributes.
Tags. Each sample is tagged with its capture phase (First, Second, Third) and its vineyard_block (B7, B9) for quick filtering in the App sidebar, in addition to being stored as regular fields.
Field descriptions. Every custom field listed below also carries a FiftyOne description attribute (set via field.description = "..."; field.save()), so the meaning of each field â especially the geometry-related bounds_/offset_/extent_ fields â is visible directly as a tooltip in the Appâs field/sidebar UI, not just in this card.
Field |
FiftyOne type |
Description |
|---|---|---|
|
|
Path to the downsampled |
|
|
|
|
|
Original source |
|
|
Flight identifier 1â10, matching Table 1 of the source paper |
|
|
Capture campaign: |
|
|
Flight date ( |
|
|
Flight time (UTC, from Table 1 of the source paper) |
|
|
Vineyard block flown: |
|
|
Reference UTM coordinates (EPSG:32629, WGS 84 / UTM zone 29N) of the vineyard block, from the source paper |
|
|
Flight altitude above ground level, in meters (20, 30, or 50) |
|
|
Programmed flight speed, verbatim from filename (e.g. |
|
|
LAS point data record format ID of the source |
|
|
CRS parsed from the source |
|
|
Number of points in the original full-resolution point cloud |
|
|
Number of points in the downsampled |
|
|
Voxel edge length (meters) used to downsample this flight to |
|
|
Size of the original source |
|
|
Minimum XYZ bounding-box coordinate of the point cloud, in the source CRS (pre-centering) |
|
|
Maximum XYZ bounding-box coordinate of the point cloud, in the source CRS (pre-centering) |
|
|
Offset subtracted from every point so the stored |
|
|
Derived footprint width, length, and height of the flightâs point cloud ( |
Dataset Creation#
Curation Rationale#
Public UAV LiDAR datasets for woody crops/vineyards are rare, despite growing interest in using LiDAR for vineyard morphology assessment, precision viticulture management, and as ground truth for satellite-derived models. This dataset was curated to fill that gap by providing a temporal series (2021â2022, three phases) of high-density, RGB-colored point clouds over two vineyard blocks at multiple flight altitudes.
Source Data#
Data Collection and Processing#
Data were collected over three flight campaigns â September 16, 2021; July 14, 2022; and September 8, 2022 â chosen to capture the vines during vegetative growth (July) and near harvest (September). Flights were flown at 20, 30, and 50 meters AGL over two commercial vineyard blocks (B7 and B9, Vitis vinifera cv. Loureiro, trained in vertical shoot positioning with spontaneous cover-crop vegetation between rows), owned by Bodegas Terras Gauda S.A. and located in Tomiño, Pontevedra, Galicia, Spain (Vineyard B7: X 517186.7, Y 4645072.3; Vineyard B9: X 516987.9, Y 4644817.7; WGS 84 / UTM zone 29N, EPSG:32629).
The UAS consisted of a DJI M300 RTK multi-rotor platform carrying a DJI Zenmuse L1 LiDAR sensor (point rate up to 240,000 pts/s single-return / 480,000 pts/s multi-return, ranging accuracy 3 cm RMS @ 100 m, 20 MP RGB mapping camera for real-time point cloud coloring). Flights were planned along the vineyard rows with DJI Pilot 2 + UgCS software at a programmed speed of 4 m/s, 50% side overlap, and 80% frontal overlap, flown autonomously with RTK positioning for precision navigation. Ten .laz (compressed LAS) point clouds with per-point RGB were produced, one per flight/altitude/block combination (see Table 1 of the source paper).
For this FiftyOne dataset, each source .laz file was parsed with laspy, and its XYZ + RGB point data converted to .pcd with open3d: a full-resolution version was written unmodified (aside from re-centering, see Dataset Structure), and a voxel-downsampled âviewâ version (~5M points) was written for interactive use in the FiftyOne App. Per-flight metadata (flight phase/date/time, vineyard block and reference coordinates, altitude, point counts, bounding box, CRS, etc.) was compiled from the source paperâs Table 1/Table 2 and from the parsed .laz headers, then attached to each FiftyOne sample as scalar fields.
Who are the source data producers?#
The data were produced by researchers at the Information Technology Group, Wageningen University & Research, as part of the H2020 FLEXIGROBOTS project, flying UAV LiDAR surveys over commercial vineyard blocks owned and operated by Bodegas Terras Gauda, S.A.
Annotations#
Annotation process#
None. This dataset consists solely of raw, unlabeled geometric point clouds with embedded per-point RGB color captured by the onboard mapping camera; no manual or automated annotation (e.g., object detection, segmentation) was performed by the dataset authors or during FiftyOne parsing.
Who are the annotators?#
Not applicable â no annotations are included.
Personal and Sensitive Information#
None. The dataset contains only aerial LiDAR point clouds and RGB imagery of vineyard vegetation and terrain; no personal or sensitive information is present.
Citation#
BibTeX:
@article{velez2023vinelidar,
title = {{VineLiDAR}: High-resolution {UAV}-{LiDAR} vineyard dataset acquired over two years in northern {Spain}},
author = {V{\'e}lez, Sergio and Ariza-Sent{\'i}s, Mar and Valente, Jo{\~a}o},
journal = {Data in Brief},
pages = {109686},
year = {2023},
issn = {2352-3409},
doi = {10.1016/j.dib.2023.109686},
url = {https://doi.org/10.1016/j.dib.2023.109686},
publisher = {Elsevier}
}
APA:
VĂ©lez, S., Ariza-SentĂs, M., & Valente, J. (2023). VineLiDAR: High-resolution UAV-LiDAR vineyard dataset acquired over two years in northern Spain. Data in Brief, 109686. https://doi.org/10.1016/j.dib.2023.109686
More Information#
Original data (Zenodo): High resolution LiDAR dataset acquired using UAV over two vineyards and two years located in âTomiñoâ, Pontevedra, Spain
Related datasets from the same vineyard site, referenced by the source paper as candidates for multimodal combination with VineLiDAR:
Ariza-SentĂs, M., VĂ©lez, S., & Valente, J. (2023). Dataset on UAV RGB videos acquired over a vineyard including bunch labels for object detection and tracking. Data in Brief, 46, 108848. https://doi.org/10.1016/j.dib.2022.108848
VĂ©lez, S., Ariza-SentĂs, M., & Valente, J. (2023). Dataset on unmanned aerial vehicle multispectral images acquired over a vineyard affected by Botrytis cinerea in northern Spain. Data in Brief, 46, 108876. https://doi.org/10.1016/j.dib.2022.108876
Dataset Card Contact#
Created by Harpreet Sahota