Tropix: Cross-Scale Tropical Tree Species Classification 🌳

Tropix is a benchmark for cross-scale knowledge distillation in tropical tree-species classification. Each labeled tree crown comes as a pair of drone images at very different scales: a crown view (≈2 cm GSD, from an orthomosaic) and a close-up (0.4 mm GSD, low-altitude drone, top-down). The task is to train a classifier on crown views that can use close-ups at training time only, since at deployment only the orthomosaic is available.

The data covers two sites, BCI (Barro Colorado Island, Panama) and TBS (Tiputini Biodiversity Station, Ecuador), with a spatially held-out train/valid/test split and, for each site, a pool of unlabeled crowns that have a close-up but no species label.

This is an anonymized version of the dataset, released for double-blind review. Authors, affiliations and the code repository are withheld and will be added after the review period.

🧭 Table of Contents

1. Overview
2. Datasets
3. Benchmark and evaluation protocol
4. Repository Structure
5. Fields
6. Data Collection & Preprocessing
7. Geospatial split
8. Usage
9. Limitations and intended use
10. License
11. Citation

📝 Overview

Each labeled row is one manually-labeled tree crown, segmented from a ≈2 cm GSD orthomosaic and paired with a 0.4 mm GSD close-up photo taken inside the same crown. The two views are non-nested: the close-up resolves fine detail the crown view lacks, and the crown view holds canopy context the close-up lacks, so neither is a pixel-level superset of the other. They are spatially, not temporally, aligned.

Crown view (orthomosaic) vs. close-up (low-altitude drone, top-down) pairs, for a common species (top, n=155) and a rare species (bottom, n=6). The crown outline (yellow) is the same polygon in both views.

📦 Datasets

Dataset	Site	Crown-view GSD	Labeled crowns	Train / Valid / Test	Species	Unlabeled pool
BCI	Barro Colorado Island, Panama	~1.87 cm/px	2,945	1,757 / 432 / 756	65	2,516
TBS	Tiputini, Ecuador	~1.86 cm/px	3,974	2,466 / 465 / 1,043	157	4,036

Species abundances are strongly long-tailed at both sites. TBS has more than twice as many species as BCI.

BCI

Crowns from the Barro Colorado Island 50-ha Forest Dynamics Plot and the surrounding drone-survey extent. Crown polygons come from CanopyRS segmentation, were labeled via Labelbox by expert botanists, independently verified (see Annotation and Quality Control), and reconciled against WCVP/GBIF taxonomy. Per-species counts are in bci_split_summary.csv.

Spatial layout of the BCI train (orange) / valid (green) / test (red) partitions. Shaded regions are the AOI boundaries defining each split's territory; outlined blobs are crown polygons. The inset shows individual crown polygons.

TBS

Same pipeline and schema as BCI (EPSG:32718, UTM 18S vs. BCI's EPSG:32617, UTM 17N, is the only schema difference).

Spatial layout of the TBS train (orange) / valid (green) / test (red) partitions. Shaded boxes are the AOI boundaries defining each split's territory (1 m-buffered crown bounding boxes, as for BCI); outlined blobs are crown polygons. Unlike BCI's large contiguous regions, TBS's splits are finely interleaved; the inset shows a region where all three partitions meet.

Unlabeled pools

Each site has a pool of crowns with a crown view and a close-up but no species label (BCI 2,516, TBS 4,036). No active-learning criterion was used to choose where close-ups were taken. The pools are filtered against the validation and test regions: every remaining crown lies more than 1 m from the valid and test AOIs, so pool crowns cannot leak evaluation context. Pool rows carry crownview_image and closeup_image; closeup_mask is null and there are no label or split columns.

🏁 Benchmark and evaluation protocol

The accompanying paper evaluates 22 teacher-based methods (labeled distillation, unlabeled-pool distillation, pseudo-labeling) and 12 no-teacher long-tailed baselines under one locked protocol. The data conventions it relies on:

• Training inputs: crown views, plus close-ups through a teacher and/or the unlabeled pool. Test inputs: crown views only.
• Metrics: Top-1 and macro-F1 (primary; every species has ≥3 test crowns, so all species enter the average), and macro-F1 within Head / Medium / Tail species groups.
• Head / Medium / Tail groups have equal training mass: species are ranked by training frequency and cut where the cumulative share of training crowns first reaches 1/3 and 2/3. Grouping by equal mass rather than equal species count keeps the Head group from swallowing most species when the distribution is long-tailed.
• Model selection: use the validation split only. Do not tune on test.

📁 Repository Structure

├── bci_spatial/
│   ├── train/*.parquet                    # 1,757 labeled crowns
│   ├── valid/*.parquet                    #   432 labeled crowns
│   ├── test/*.parquet                     #   756 labeled crowns
│   ├── unlabeled/*.parquet                # 2,516 unlabeled crowns (crown view + close-up, no label)
│   ├── bci_split_summary.csv              # per-species total/train/valid/test counts
│   ├── bci_{train,valid,test}_aoi.gpkg    # AOI boundary polygons defining each split region
│   ├── bci_{train,valid,test}_polygons.gpkg  # per-split crown polygons
│   └── bci_all_polygons.gpkg              # all labeled crown polygons, any split
├── tbs_spatial/
│   └── (identical layout, tbs_ prefix; unlabeled/ holds 4,036 crowns)
├── figures/
└── README.md

The .gpkg files are included so the spatial holdout can be independently verified: they are the actual AOI/polygon geometries used to assign each crown to a split.

🗂️ Fields

Labeled parquet rows have the columns below. All scalar fields are stored as strings (cast numeric ones yourself); image fields are {bytes, path} structs with path always None.

Field	Stored type	Description
canopyrs_object_id	string	Stable crown ID (unique within site; disjoint across splits)
canonical_name	string	Label. GBIF canonical species name
scientific_name	string	Full scientific name with author
genus, family	string	Taxonomic hierarchy
rank	string	Taxon rank (e.g. SPECIES)
area	string→float	Crown polygon area, m²
gbif_id	string→float	GBIF taxon key
wcvp_name	string	WCVP-reconciled name
wcvp_url	string	Link to this species' WCVP/POWO taxon page
site	string	"bci" or "tbs"
polygon	string (WKT)	Crown geometry, in crs coordinates
crs	string	"EPSG:32617" (BCI) or "EPSG:32718" (TBS)
transform	string (JSON list)	Affine (6 floats) georeferencing the crown-view raster
source_orthomosaic	string	Drone-survey block the crown's centroid falls in (each site was flown as several overlapping sub-area missions)
annotation_mask_coverage	string→float	% of crown bounding box covered by the mask
crownview_image	struct{bytes,path}	JPEG, RGB, native resolution (~1000×1000, variable). Aerial crop over the crown polygon, with surrounding canopy context
closeup_image	struct{bytes,path}	JPEG, RGB, native resolution (~4000×3000). Low-altitude drone, top-down
closeup_mask	struct{bytes,path}	PNG, mode L, matches close-up size. Crown mask for the close-up

Unlabeled parquet rows have only: canopyrs_object_id, area, site, polygon, crs, transform, crownview_image, closeup_image, and closeup_mask (always null).

Images are native resolution, not resized.

Taxonomy versions: gbif_id resolves against the GBIF Backbone Taxonomy; wcvp_name/wcvp_url against WCVP version 16 (extracted 2026-06-04). A few taxa may shift accepted name or classification if those databases are later updated.

🛰️ Data Collection & Preprocessing

1. Orthomosaic (crown views)

RGB orthomosaics at ≈2 cm GSD (1.87 cm BCI, 1.86 cm TBS). This is the source raster crown views are cropped from and the input to CanopyRS segmentation.

2. Close-up capture

Close-ups are drone-captured top-down photographs at 0.4 mm GSD, following an automated close-up mission-planning workflow (Laliberté et al.). The close-up location within a crown was chosen randomly but always near the tree center.

3. Segmentation

CanopyRS produces crown segmentation polygons (geometry only, no species label and no close-up link) from the orthomosaic.

4. Annotation and Quality Control

• Waypoint join: the waypoint layer (one point per close-up, each carrying a zoom_url) is joined spatially onto the segmentation polygons: a waypoint attaches its zoom_url to the polygon containing it.
• Close-up mask curation: close-ups were annotated in Labelbox with per-pixel binary masks. Duplicate masks with the same species label on one close-up were merged. Close-ups whose masks covered multiple species were manually reviewed (n=83 BCI, n=165 TBS). When a close-up had several single-species mask candidates, the one with the highest canopy coverage was kept if it exceeded 50% of the image area, unless the coverage gap between candidates was under 20% (then manual review).
• Labeling join: validated species labels are joined by attribute match (zoom_url ↔ Labelbox data_row.row_data). When several waypoints fell inside one polygon, one label was kept if all agreed (or only one carried a valid label); residual ambiguous cases were resolved manually (n=3 BCI, n=19 TBS).
• Independent verification (BCI): every labeled crown is checked by a second botanist (confirm / reject, with optional corrected species and note); flagged crowns go to a third reviewer for adjudication, and any case still unresolved goes through a final round. Names are reconciled for synonymy under current taxonomy. TBS has no equivalent second pass.

5. Taxonomy matching

Species names are resolved against WCVP and GBIF.

🗺️ Geospatial split

Why. A species-stratified random split treats crowns as exchangeable. For spatially referenced imagery that fails: a train crown and a test crown a few meters apart share background, illumination and even overlapping pixels, so a model can score well from local context rather than species recognition. A valid split must also enforce spatial disjointness.

Spatial cliques. Each crown polygon is expanded by a 1 m canopy-context buffer, and overlapping bounding boxes are merged into disjoint spatial cliques with union-find. All crowns of a clique go to the same partition.

Allocation by integer programming. Cliques are allocated to train / valid / test with a mixed-integer linear program (CBC) that:

• keeps overall proportions close to the 60 / 15 / 25 targets,
• guarantees every evaluated species at least 1 training crown and 3 test crowns, and
• minimizes how far any species' share in a split deviates from its target.

Species whose crowns cannot satisfy these floors are dropped, fewest first. This is a structural consequence of requiring spatial validity and per-class statistical reliability together.

Resulting split. BCI: 1,757 / 432 / 756 crowns (59.7 / 14.7 / 25.7%). TBS: 2,466 / 465 / 1,043 (62.1 / 11.7 / 26.2%). The median test crown lies 27.6 m (BCI) and 17.6 m (TBS) from the nearest training crown.

Site	Raw species	≥ 4 samples	Published (post geospatial-solve)
BCI	97	71	65
TBS	313	159	157

(BCI's raw count is after the independent verification; TBS's is the direct Labelbox output.)

BCI: continuity with the 50-ha plot. To stay comparable with prior work on the BCI 50-ha plot, the previously established partition within the plot is kept unchanged and the solver is applied only to the surrounding extent.

Each labeled polygon is assigned to a split by centroid-in-AOI (priority test > valid > train). The AOI and per-split polygon .gpkg files are published so the split can be verified independently.

Conversion to ML-ready data. For each labeled, split-assigned polygon, a crown-view image is cropped from the raster at the polygon bounding box plus a pixel buffer (native resolution), and the close-up and its mask are packed alongside. A small number of polygons exported twice under the same Labelbox label were deduplicated by canopyrs_object_id.

💻 Usage

from datasets import load_dataset, Image

# Labeled data (stream to avoid downloading the full ~5 GB train shard)
ds = load_dataset("anon567473/tropix", name="bci_spatial", streaming=True)
example = next(iter(ds["train"]))
crown_img = Image().decode_example(example["crownview_image"])    # PIL, native resolution
closeup_img = Image().decode_example(example["closeup_image"])    # PIL, native resolution
species = example["canonical_name"]

# Unlabeled pool (no labels)
pool = load_dataset("anon567473/tropix", name="bci_spatial_unlabeled", split="unlabeled", streaming=True)

Drop streaming=True to download and cache locally. Code for the data pipeline and benchmark: anonymous code link to be added.

⚠️ Limitations and intended use

• Long tail: many species have very few crowns; the 3-test-crown floor means rare-species per-class metrics rest on 3–5 individuals. Report macro-F1 and Head/Medium/Tail breakdowns, not only Top-1.
• Geographic scope: two sites, two forests. Results may not transfer to other regions, sensors or seasons.
• Label quality: BCI labels have a second-botanist verification pass; TBS labels do not, and residual labeling errors are possible.
• Non-nested views: the close-up does not contain the crown view's context (and vice versa), so methods assuming co-registered or nested views may not apply.
• Intended use: research on classifying species from crown views, long-tailed and semi-supervised learning, and cross-scale distillation. Not intended for individual-tree identification, enforcement, or any decision without field validation.

⚖️ License

Released under CC BY 4.0.

📚 Citation

Anonymous. XTropic: Benchmarking Cross-Scale Knowledge Distillation for
Tropical Tree Species Classification. Under review.

Full citation, authors and acknowledgments will be added after the review period.