---
license: mit
library_name: sklearn
tags:
  - remote-sensing
  - tree-canopy
  - sentinel-2
  - philippines
  - metro-manila
  - civic-technology
pipeline_tag: tabular-classification
---

# Leaves.PH canopy classifier

The published per-pixel canopy model behind [leaves.ph](https://leaves.ph), an open, reproducible tree-cover map of Metro Manila (17 NCR cities plus Pateros), 2019 to 2026.

A `HistGradientBoostingClassifier` (scikit-learn) that labels each 30 m Sentinel-2 pixel as canopy or not, trained on **656 hand-labeled high-resolution pixels**. It is the source of every figure on the site: the map, the per-LGU and per-barangay series, and the headline NCR canopy percentage (about 9 to 10 percent).

## Features (10)

`ndvi`, `dw` (Dynamic World tree probability), `meta_h` (Meta v2 1 m canopy height), `esatree` (ESA WorldCover tree class), the raw Sentinel-2 bands `red` / `nir` / `green` / `blue`, plus `gndvi` and `nir_red`. Decision threshold 0.5, calibrated so 2021 matches the 10.1 percent human-truth canopy.

## Evaluation

Scored against the 656 manual labels under region-grouped out-of-fold cross-validation with post-stratified population weighting:

| Model | Precision | Recall | F1 | IoU |
|---|---|---|---|---|
| This classifier (10 features) | 0.77 | 0.79 | **0.78** | **0.64** |
| Four-feature model (no spectral bands) | - | - | 0.75 | - |
| NDVI > 0.62 baseline | 0.69 | 0.67 | 0.68 | 0.52 |

The raw green and blue bands are what lift it over the baseline: they let it reject high-NDVI grass and scrub the NDVI threshold over-called (precision 0.67 to 0.77), and it removes the year-to-year sawtooth the fixed threshold produced. A CLIP ViT-L/14 embedding was tested as a feature and did not help, so it was dropped.

## Intended use and limits

- It is a tree-canopy estimate with a known grass/scrub margin, not a per-tree census and not a parcel-level land-use tool.
- "Canopy" is a dense-vegetation proxy. Roughly half the flagged area is verified tree canopy above 5 m.
- The threshold is calibrated to one epoch (Meta's source imagery is mostly 2018 to 2020), so read the per-year values as annual cross-sectional snapshots, not a validated change series.

## Files

- `canopy_clf.joblib` - the trained classifier
- `canopy_clf_meta.json` - features, threshold, training size
- `master_labels.csv` - the 656 gold labels
- `model_comparison.json` - the ablation it was selected from
- `RESULTS.md` - the build writeup

## Links and citation

- Code and pipeline: https://github.com/xmpuspus/leaves-ph (MIT)
- Data products on HuggingFace: https://huggingface.co/datasets/xmpuspus/leaves-ph (CC-BY-4.0)
- Methodology: https://leaves.ph/methodology
- Cite: Leaves.PH (2026, v0.7.0). https://doi.org/10.5281/zenodo.20470306