🌲 Forest and trees v.2026-09-10 — per-location benchmark
This page details the validation of the 🌲 Forest and trees v.2026-09-10 segmentation model on 8 areas of interest (AOI), compared against the previous version v.2025-06-14. For each AOI the two prediction masks are shown side by side; click any image to open it full size, and use the ← / → arrow keys to browse between them.
All metrics are area-based: IoU is the intersection-over-union of the predicted and ground-truth vegetation masks, and F1 / Precision / Recall are computed on the overlapping mask area. The segmentation output is unchanged from v.2026-07-03 (the v.2026-09-10 update refines crown extraction), so these per-location figures carry over. Segmentation evaluation run: 2026-07-03.
Mask colour legend: v.2026-09-10 · v.2025-06-14
Filipinas — Balanga
Modelo |
pagaré |
F1 |
Precisión |
Recordar |
|---|---|---|---|---|
v.2026-09-10 |
0,806 |
0,892 |
0.864 |
0,923 |
v.2025-06-14 |
0.593 |
0.745 |
0,914 |
0,628 |
España — Velilla de San Antonio
Modelo |
pagaré |
F1 |
Precisión |
Recordar |
|---|---|---|---|---|
v.2026-09-10 |
0,784 |
0,879 |
0,894 |
0,864 |
v.2025-06-14 |
0.008 |
0,015 |
0,989 |
0.008 |
España — Madrid
Modelo |
pagaré |
F1 |
Precisión |
Recordar |
|---|---|---|---|---|
v.2026-09-10 |
0,813 |
0,897 |
0.880 |
0,915 |
v.2025-06-14 |
0,447 |
0,618 |
0,973 |
0.453 |
Italia — Crotona
Modelo |
pagaré |
F1 |
Precisión |
Recordar |
|---|---|---|---|---|
v.2026-09-10 |
0,862 |
0,926 |
0.944 |
0,909 |
v.2025-06-14 |
0.803 |
0.890 |
0,988 |
0.811 |
España — Rusia
Modelo |
pagaré |
F1 |
Precisión |
Recordar |
|---|---|---|---|---|
v.2026-09-10 |
0,617 |
0,763 |
0,699 |
0,841 |
v.2025-06-14 |
0.057 |
0.108 |
0,964 |
0.057 |
España — Cuenca
Modelo |
pagaré |
F1 |
Precisión |
Recordar |
|---|---|---|---|---|
v.2026-09-10 |
0,763 |
0,865 |
0,859 |
0,872 |
v.2025-06-14 |
0,459 |
0,629 |
0,873 |
0,492 |
Argentina — La Banda
Modelo |
pagaré |
F1 |
Precisión |
Recordar |
|---|---|---|---|---|
v.2026-09-10 |
0,576 |
0,731 |
0,653 |
0,829 |
v.2025-06-14 |
0.233 |
0,378 |
0,958 |
0.235 |
Uzbekistán — Taskent
Modelo |
pagaré |
F1 |
Precisión |
Recordar |
|---|---|---|---|---|
v.2026-09-10 |
0,730 |
0,844 |
0.905 |
0,791 |
v.2025-06-14 |
0.568 |
0,724 |
0,981 |
0.574 |
Resumen
v.2026-09-10 improves substantially over the previous v.2025-06-14 across all 8 AOIs. Mean area-based F1 rises from 0.513 to 0.850 and IoU from 0.396 to 0.744, driven by a large recall gain (0.407 → 0.868). The previous version was highly conservative — mean precision 0.955, but it missed most vegetation and collapsed on several AOIs (e.g. Velilla de San Antonio F1 0.015, Rus F1 0.108, La Banda F1 0.378). v.2026-09-10 keeps precision high (0.837) while recovering the missed canopy, and leads on F1 in every location.