🌲 Forest and trees v.2026-09-10 — tree crown benchmark
This page details the validation of tree crown polygons produced by the 🌲 Forest and trees v.2026-09-10 model on 7 areas of interest (AOI), compared against the previous version v.2026-07-03. For each AOI the two crown-mask results are shown side by side; click any image to open it full size, and use the ← / → arrow keys to browse between them.
Metrics are area-based on the crown masks: IoU is the intersection-over-union of the predicted and ground-truth crown masks, and F1 / Precision / Recall are computed on the overlapping mask area. Crowns is the number of detected crown polygons; the ground-truth crown count and the relative count error are reported below each AOI (see also the Tree count accuracy table on the model page). Evaluation run: 2026-10-01.
Mask colour legend: v.2026-09-10 · v.2026-07-03
Italy — Crotone
Model |
IoU |
F1 |
Precision |
Recall |
Crowns |
|---|---|---|---|---|---|
v.2026-09-10 |
0.842 |
0.914 |
0.930 |
0.898 |
254 |
v.2026-07-03 |
0.825 |
0.904 |
0.925 |
0.884 |
254 |
Ground-truth crowns: 327. Relative count error — v.2026-09-10: 22.3%; v.2026-07-03: 22.3%.
Argentina — La Banda
Model |
IoU |
F1 |
Precision |
Recall |
Crowns |
|---|---|---|---|---|---|
v.2026-09-10 |
0.624 |
0.769 |
0.700 |
0.852 |
111 |
v.2026-07-03 |
0.575 |
0.730 |
0.691 |
0.774 |
83 |
Ground-truth crowns: 125. Relative count error — v.2026-09-10: 11.2%; v.2026-07-03: 33.6%.
Spain — Rus
Model |
IoU |
F1 |
Precision |
Recall |
Crowns |
|---|---|---|---|---|---|
v.2026-09-10 |
0.689 |
0.816 |
0.759 |
0.883 |
239 |
v.2026-07-03 |
0.567 |
0.723 |
0.671 |
0.785 |
249 |
Ground-truth crowns: 258. Relative count error — v.2026-09-10: 7.4%; v.2026-07-03: 3.5%.
Philippines — Balanga
Model |
IoU |
F1 |
Precision |
Recall |
Crowns |
|---|---|---|---|---|---|
v.2026-09-10 |
0.780 |
0.876 |
0.849 |
0.904 |
309 |
v.2026-07-03 |
0.782 |
0.878 |
0.860 |
0.896 |
220 |
Ground-truth crowns: 193. Relative count error — v.2026-09-10: 60.1%; v.2026-07-03: 14.0%.
Spain — Cuenca
Model |
IoU |
F1 |
Precision |
Recall |
Crowns |
|---|---|---|---|---|---|
v.2026-09-10 |
0.732 |
0.846 |
0.875 |
0.818 |
433 |
v.2026-07-03 |
0.728 |
0.842 |
0.868 |
0.818 |
409 |
Ground-truth crowns: 321. Relative count error — v.2026-09-10: 34.9%; v.2026-07-03: 27.4%.
Spain — Velilla de San Antonio
Model |
IoU |
F1 |
Precision |
Recall |
Crowns |
|---|---|---|---|---|---|
v.2026-09-10 |
0.694 |
0.819 |
0.855 |
0.786 |
1,134 |
v.2026-07-03 |
0.667 |
0.800 |
0.831 |
0.771 |
1,070 |
Ground-truth crowns: 1,212. Relative count error — v.2026-09-10: 6.4%; v.2026-07-03: 11.7%.
Spain — Madrid
Model |
IoU |
F1 |
Precision |
Recall |
Crowns |
|---|---|---|---|---|---|
v.2026-09-10 |
0.823 |
0.903 |
0.907 |
0.899 |
398 |
v.2026-07-03 |
0.801 |
0.889 |
0.879 |
0.900 |
368 |
Ground-truth crowns: 478. Relative count error — v.2026-09-10: 16.7%; v.2026-07-03: 23.0%.
Summary
For tree crown polygons, v.2026-09-10 improves on v.2026-07-03 across the 7 AOIs: mean area-based F1 rises from 0.824 to 0.849 and IoU from 0.706 to 0.741, with gains in both precision (0.818 → 0.839) and recall (0.833 → 0.863). It leads on F1 in 6 of the 7 AOIs and is on par at Balanga (0.876 vs 0.878).
On tree count accuracy the picture is mixed. Where the previous model under-detected, v.2026-09-10 is markedly closer to the true crown count (La Banda 33.6% → 11.2%, Madrid 23.0% → 16.7%, Velilla de San Antonio 11.7% → 6.4%). But in two dense-canopy AOIs it over-segments the canopy, producing more crowns than the ground truth (Balanga 60.1% vs 14.0%, Cuenca 34.9% vs 27.4%), which raises its mean relative count error to 22.7% versus 19.4% for v.2026-07-03 across all 7 AOIs. Mask-overlap accuracy (IoU / F1) still favours v.2026-09-10 in those areas; reducing over-segmentation in dense canopy is the focus for the next iteration.