Coarse woody debris inventory from UAV imagery

Code for a site-wide coarse woody debris inventory of a Hurricane Helene windthrow site in the Georgia Coastal Plain, built from UAV orthomosaic imagery and instance segmentation.

The pipeline detects fallen woody material, extracts a centreline for each instance, samples perpendicular widths along it, integrates conical frusta to estimate volume, and classifies each object as a full stem, a fragment, or one of four exclusion categories.

The imagery and all georeferenced products are not released. The study site is privately owned land. See docs/DATA.md.

Headline results

Quantity	Value
Imaged area	478.7345 ha
Total detections	23,603
Inventory objects	21,205
Inventory volume	14,237.0 m3
Inventory density	29.74 m3/ha, 44.3 objects/ha
Full stems	5,240 objects, 8,041.1 m3
Fragments	15,965 objects, 6,195.9 m3

Full stems are 24.7 percent of counted objects and 56.5 percent of volume.

Volumes are estimator output. They have never been calibrated against field measurement. The object count is a floor rather than a ceiling, because recall of 0.437 dominates any inflation from double counting.

Every number the code must reproduce is listed in docs/NUMBERS.md.

Detection performance

42-image validation split, scored with pycocotools at maxDets 700, score threshold 0.001, full-resolution masks, no ground-truth size filter.

Model	Box AP50	Box AP	Mask AP50	Mask AP
Mask R-CNN	40.8	19.1	27.1	6.9
YOLOv8s-seg	43.0	21.9	24.9	5.8

Both models were originally scored by their own frameworks, which do not agree. Rescoring both through one evaluator removed an apparent mask AP50 advantage entirely. The five protocol differences behind that are documented in src/cwd/evaluate.py.

YOLOv8s-seg is the inventory model, chosen on operational grounds rather than an accuracy claim.

Layout

configs/
  thresholds.json     classification rules, the single source of truth
  maskrcnn.yaml       Mask R-CNN architecture and solver
  yolov8s_seg.yaml    YOLOv8s-seg training arguments
src/cwd/
  paths.py            locate the data root by content, never by hardcoded path
  volume_pipeline.py  per-object measurement and frustum volume
  classify.py         inventory and exclusion labels, with a self-test
  evaluate.py         unified pycocotools scoring for both models
  figures.py          figure style and save helpers
docs/
  DATA.md             what is released, what is not, how to request access
  NUMBERS.md          every locked number the code must reproduce
  REPRODUCE.md        pipeline order and the traps that have cost time
  FIGURES.md          figure conventions and per-figure inputs
scripts/
  preflight.sh        pre-commit scan, seven failure classes
  fetch_data.py       download the Zenodo deposit

Verifying the classification rules

This runs without the imagery. It re-derives every label from a released measurement table and asserts the published per-class counts.

PYTHONPATH=src python -m cwd.classify path/to/cwd_classified_site.csv

Expected: 23,603 rows, agreement 1.0000, and the six counts in docs/NUMBERS.md.

The order of the threshold tests is definitional. Applying the length floor before the width floor yields the same inventory count but splits the exclusions 1,178 / 461 instead of 855 / 784. The assertion is what pins the order down.

Classification rules

Thresholds are in configs/thresholds.json and match Table 6 of the paper.

Tests apply in this order:

1. width_corr_m below 0.15 m, three pixels, excluded
2. length_m below 1.37 m, excluded
3. elongation below 3.0, excluded
4. touches a tile edge, excluded, because length on a truncated mask is unreliable
5. full stem if length is at least 6.78 m, elongation at least 12.0, and corrected width between 0.25 and 1.50 m
6. otherwise a fragment

Width comparisons use width_corr_m, which is width_mean_m / 1.19. Predicted masks are about 19 percent wider than hand-annotated ones at the median. The correction is applied to classification only, never to reported volumes.

Environment

PyTorch 2.11.0+cu128, Detectron2 0.6, Ultralytics 8.4.101, Python 3.12.13, CUDA 12.8. See requirements.txt. Detectron2 has no PyPI release and must be built from source.

Citation

See CITATION.cff. Code is MIT licensed; the data deposit is CC BY 4.0. Requests for the full dataset go to the corresponding author.

Authors

Dang Hoang, Rishab Subramaniyan, Owen Norman, Joey Hattan, Roger C. Lowe III. AI@UGA and the Warnell School of Forestry and Natural Resources, University of Georgia; Brown University.