Leaves.PH

SECTION 5 · VALIDATION GALLERY · DETECTION MODEL · t=0·60

Per-LGU visual validation of the detection model.

One panel per city. Each lays the model's canopy reading over the satellite image so you can see, block by block, where it agrees with the simpler NDVI rule and where the two differ. The table below ranks the 17 LGUs by how much canopy the model reads.

NDVI v0 baseline model-confirmed canopy model differs from NDVI (a difference, not a verified correction)

Under the hood, the detection model is a CLIP ViT-Large/14 head with a gradient-boosted regression onto Meta's 1 m canopy fraction; on held-out locations it reaches R² 0.83 to 0.86 under grouped cross-validation against Meta's 1 m reference (its calibration target, not independent ground truth), and is in optimization toward a first release. Panels are sorted by the model's canopy percentage at threshold t = 0.60.

№ LGU NDVI v0 Model %
01 Quezon City 19.53% 18.74%
02 Mandaluyong 10.59% 12.11%
03 Taguig 8.24% 10.22%
04 Makati 7.89% 8.61%
05 Valenzuela 7.64% 7.56%
06 Marikina 8.43% 7.51%
07 Caloocan 8.79% 5.79%
08 Malabon 5.28% 5.60%
09 Pasig 5.60% 3.84%
10 Muntinlupa 3.61% 3.51%
11 Las Pinas 4.58% 3.26%
12 Paranaque 3.12% 2.12%
13 Pasay 2.50% 2.10%
14 Pateros 4.68% 0.84%
15 Manila 0.69% 0.52%
16 Navotas 0.39% 0.45%
17 San Juan 2.31% 0.21%
Validation panel for Quezon City: S-2 RGB, NDVI v0 baseline (orange), detection model shared (green), detection model NEW additions (yellow), stats panel.
VAL · Quezon City · NDVI 19.53% / model 18.74% / Δ -0.79 pp t = 0·60
Validation panel for Mandaluyong: S-2 RGB, NDVI v0 baseline (orange), detection model shared (green), detection model NEW additions (yellow), stats panel.
VAL · Mandaluyong · NDVI 10.59% / model 12.11% / Δ +1.51 pp t = 0·60
Validation panel for Taguig: S-2 RGB, NDVI v0 baseline (orange), detection model shared (green), detection model NEW additions (yellow), stats panel.
VAL · Taguig · NDVI 8.24% / model 10.22% / Δ +1.98 pp t = 0·60
Validation panel for Makati: S-2 RGB, NDVI v0 baseline (orange), detection model shared (green), detection model NEW additions (yellow), stats panel.
VAL · Makati · NDVI 7.89% / model 8.61% / Δ +0.71 pp t = 0·60
Validation panel for Valenzuela: S-2 RGB, NDVI v0 baseline (orange), detection model shared (green), detection model NEW additions (yellow), stats panel.
VAL · Valenzuela · NDVI 7.64% / model 7.56% / Δ -0.09 pp t = 0·60
Validation panel for Marikina: S-2 RGB, NDVI v0 baseline (orange), detection model shared (green), detection model NEW additions (yellow), stats panel.
VAL · Marikina · NDVI 8.43% / model 7.51% / Δ -0.91 pp t = 0·60
Validation panel for Caloocan: S-2 RGB, NDVI v0 baseline (orange), detection model shared (green), detection model NEW additions (yellow), stats panel.
VAL · Caloocan · NDVI 8.79% / model 5.79% / Δ -3.00 pp t = 0·60
Validation panel for Malabon: S-2 RGB, NDVI v0 baseline (orange), detection model shared (green), detection model NEW additions (yellow), stats panel.
VAL · Malabon · NDVI 5.28% / model 5.60% / Δ +0.32 pp t = 0·60
Validation panel for Pasig: S-2 RGB, NDVI v0 baseline (orange), detection model shared (green), detection model NEW additions (yellow), stats panel.
VAL · Pasig · NDVI 5.60% / model 3.84% / Δ -1.76 pp t = 0·60
Validation panel for Muntinlupa: S-2 RGB, NDVI v0 baseline (orange), detection model shared (green), detection model NEW additions (yellow), stats panel.
VAL · Muntinlupa · NDVI 3.61% / model 3.51% / Δ -0.10 pp t = 0·60
Validation panel for Las Pinas: S-2 RGB, NDVI v0 baseline (orange), detection model shared (green), detection model NEW additions (yellow), stats panel.
VAL · Las Pinas · NDVI 4.58% / model 3.26% / Δ -1.32 pp t = 0·60
Validation panel for Paranaque: S-2 RGB, NDVI v0 baseline (orange), detection model shared (green), detection model NEW additions (yellow), stats panel.
VAL · Paranaque · NDVI 3.12% / model 2.12% / Δ -1.00 pp t = 0·60
Validation panel for Pasay: S-2 RGB, NDVI v0 baseline (orange), detection model shared (green), detection model NEW additions (yellow), stats panel.
VAL · Pasay · NDVI 2.50% / model 2.10% / Δ -0.40 pp t = 0·60
Validation panel for Pateros: S-2 RGB, NDVI v0 baseline (orange), detection model shared (green), detection model NEW additions (yellow), stats panel.
VAL · Pateros · NDVI 4.68% / model 0.84% / Δ -3.84 pp t = 0·60
Validation panel for Manila: S-2 RGB, NDVI v0 baseline (orange), detection model shared (green), detection model NEW additions (yellow), stats panel.
VAL · Manila · NDVI 0.69% / model 0.52% / Δ -0.17 pp t = 0·60
Validation panel for Navotas: S-2 RGB, NDVI v0 baseline (orange), detection model shared (green), detection model NEW additions (yellow), stats panel.
VAL · Navotas · NDVI 0.39% / model 0.45% / Δ +0.06 pp t = 0·60
Validation panel for San Juan: S-2 RGB, NDVI v0 baseline (orange), detection model shared (green), detection model NEW additions (yellow), stats panel.
VAL · San Juan · NDVI 2.31% / model 0.21% / Δ -2.10 pp t = 0·60

The scan scripts that build these panels are in the public repository (the per-year probability rasters and threshold sweeps are research-track and regenerable, not committed). Per-LGU table + full pipeline in /methodology.