# R2: HJA TV081 LiDAR stem map (2008), benchmarked with NativeGIS

Draft, not published. Run 2026-10-30. NativeGIS 0.2.0.

## What was reproduced

**Source:** HJ Andrews Experimental Forest, dataset TV081, "Forest metrics derived from the 2008 LiDAR point clouds". Spies, T.A. (PI); EDI knb-lter-and.5316.4. This item is the stem map `tv08117`, made by T. Valentine in 2014.
- **Published method (FGDC metadata):** FUSION `CanopyModel` 0.5 m canopy model, then TreeVaW (Popescu) with window thresholds between 3 and 23 m. Crown radius comes from height-to-crown equations from the literature for 10–60 m trees, and a self-fitted relationship for taller trees. Trees ≥10 m only.
- **Published number:** **1,328,242 trees ≥10 m** across the 2008 HJA LiDAR area, each with a height and a crown radius.

**Data:**
- Reference: `tv08117.zip` (37,235,742 bytes, sha256 98f5f73b…494b02). Fetched anonymously from https://andrewsforest.oregonstate.edu/sites/default/files/lter/data/studies/spatialdata/tv08117.zip and kept under `samples/tv081/`. Use constraints: "no legal constraints", acknowledgement requested. HJA's data policy asks users to identify themselves; this URL does not enforce it.
- LiDAR: the same 2008 flight (Watershed Sciences, 10–11 Aug 2008), from OpenTopography OTLAS.082011.26910.1, tile `ot_HJ_Andrews_bin_100.laz`. That is one 750 m × 750 m window, E 565350–566100, N 4900150–4900900, EPSG:26910, 9.77 pts/m².

**What a full reproduction needs:** all 189 tiles (4.85 GB) and a tiled whole-forest run. The tiled run exists only as a driver script (`tools/scaleout`), not as an MCP tool. This run is therefore a **one-window benchmark**: the same flight and the same CHM resolution, but NativeGIS's own detector rather than TreeVaW.

## Method (step numbers are those in `run.json`)

1. Open the stem map [1]: 1,328,242 points, EPSG:26910 identified from its CRS name. Open the LAZ tile [3] and run QA [4]: 5,493,081 points, 0.35 ground pts/m².
2. Select the reference trees in the window [5]: **16,624**.
3. Ground: auto mode kept the delivered class 2 (0.35 /m² ≥ 0.25) [6]. DTM: 1 m TIN [8]. CHM: **0.5 m** pit-free [10], the same grid size as TV081.
4. Trees [12]: NativeGIS `gis_lidar_trees`. This is a Popescu & Wynne (2004) variable-window local-maximum detector, the same family as TreeVaW.
   - Settings: **minHeight 10 m**, **wsMin 3 m** (TV081's lower window bound), and the PNW default crown equation `ws = max(3, 0.6·(2 + 0.18·h))`.
   - CHM smoothing: Gaussian, σ = 1 cell (the default).
   - Result: 10,261 trees in the tile.
5. Match against the reference [14] with `gis_compare` trees mode: one-to-one Hungarian matching within 3 m, inside a **20 m inset** of the window (E 565370–566080, N 4900170–4900880). The inset keeps trees whose crowns are cut by the tile edge out of the score.
6. Sensitivity run: smoothing σ = 0 [15], scored the same way [17]. TV081 does not say whether TreeVaW smoothed the CHM. **No other parameter was varied, and nothing was fitted to the reference.**

## Results (inset window, 50.4 ha; n reference = 14,997 trees ≥10 m)

| | Default (σ = 1) [14] | No smoothing (σ = 0) [17] |
|---|---:|---:|
| Detected | 9,188 | 11,462 |
| Matched | 9,054 | 11,033 |
| Recall (1 − omission) | 60.4% | 73.6% |
| Precision (1 − commission) | 98.5% | 96.3% |
| F-score | 0.749 | 0.834 |
| Height bias, NativeGIS − TV081 (matched) | −0.43 m | −0.18 m |
| Height RMSE | 0.91 m | 0.74 m |
| Mean position offset | 0.48 m | 0.52 m |

**By reference height class**, with smoothing off [17]:

| Class | Reference trees | Omission | Height bias |
|---|---:|---:|---:|
| 10–20 m | 10,195 | 34.2% | −0.11 m |
| 20–30 m | 1,795 | 15.0% | −0.22 m |
| 30–40 m | 1,436 | 7.5% | −0.24 m |
| 40–50 m | 580 | 9.0% | −0.38 m |
| 50–60 m | 344 | 7.6% | −0.43 m |
| 60–70 m | 522 | 4.2% | −0.34 m |
| 70+ m | 125 | 1.6% | −0.40 m |

## Reading it honestly

- **The published total is confirmed from the published file:** 1,328,242 features. The count over the whole forest is not reproduced here; that needs the whole flight.
- **Where the canopy is tall, NativeGIS and TV081 agree closely.** For trees ≥30 m, omission is 2–9%. Heights agree to within about 0.2–0.4 m, and positions to about 0.5 m. Both methods read the same 0.5 m canopy surface from the same points, so this agreement is expected. It is still a useful check that the CHM step is equivalent.
- **The difference is the small trees.** 68% of the reference trees in this window are 10–20 m tall. On `trees_vs_tv081.svg`, the omitted trees (red) cluster in the dense young stand in the north-east quarter. In the 10–20 m class, TV081 has 10,195 tops and NativeGIS detects 6,978 even without smoothing, so TV081 has 1.46× as many.
  - We cannot say which is closer to the ground truth. **No field stem map is public for this window.** TV010 plot data sits behind the HJA registration form.
  - A larger count is not automatically better: TreeVaW at 0.5 m can split one crown into several tops.
- **Smoothing explains part of the gap.** Turning it off raises recall by 13 points for a 2-point loss in precision. The remaining difference is the window-size equation:
  - TV081 used literature equations it does not name, plus a self-fitted one for trees >60 m.
  - NativeGIS has no maximum window (TV081 capped at 23 m) and no way to supply a custom height-to-crown equation (gap ticket).
- **Not done:** tuning NativeGIS to maximise agreement with TV081. That would turn a benchmark into a fit.

## Replay

- `run.json`: 18 steps, relative to `NATIVEGIS_DATA`, sha256 6c2c8eb9…738f64.
- Replayed in a fresh session into `replay-1/`: **18 of 18 identical**, 0 differ, 0 failed.
- Cited report: `report.html`, written by `gis_report`, which cites 8 steps (sha256 534f182e…b4be).
