EO4Change Group 4 — Post-Fire Forest Recovery Monitor

Sentinel-2 + Landsat 8/9 pipeline that maps per-pixel forest recovery after wildfires using a Relative Recovery Indicator (RRI) and Land Surface Temperature. Built for the DTU Space EO4Change course, June 2026.

What it does

Builds cloud-free seasonal composites from Sentinel-2 for three periods: pre-fire, post-fire (fire scar), and recent.
Computes a per-pixel RRI — how much of the pre-fire NDVI has been restored:
```
RRI = (recent_NDVI − postfire_NDVI) / (prefire_NDVI − postfire_NDVI)
```
Classifies pixels into a recovery class: recovering well / recovering weakly / not recovering / outside assessment.
Generates time-series plots (RRI, NDVI, NDRE, NBR) and a pixel-count histogram by class.
Exports an interactive HTML map with all layers, a dynamic legend, and a Chart.js time-series panel.
Optionally adds a Land Surface Temperature layer (Landsat 8/9 Collection 2 L2SP).

Study sites

Region	Config	Reference event
Rugballegård Skov, Denmark	`data/config_DK.yaml`	Tree planting April 2022
Pedrógão Grande, Portugal	`data/config_pt.yaml`	Wildfire June 2017

Setup

# Clone and enter the repo
git clone https://github.com/leahiebel/EO4Change-Tree-Replanting.git
cd EO4Change-Tree-Replanting

# Create and activate a virtual environment
python -m venv ../venv
source ../venv/bin/activate          # macOS/Linux
# ../venv/Scripts/activate           # Windows

# Install dependencies
pip install earthengine-api folium pyyaml pandas matplotlib python-dateutil

# Authenticate with Google Earth Engine (one-time)
python -c "import ee; ee.Authenticate()"

Running

# Denmark — default config
python src/gee.py

# Portugal
python src/gee.py --config config_pt.yaml

# Second run is fast: time-series loaded from cached CSV automatically
python src/gee.py --config config_pt.yaml

# Force re-extraction from GEE (discard the cache)
python src/gee.py --config config_pt.yaml --force-ts

# Skip time-series entirely (map only, no plots)
python src/gee.py --config config_pt.yaml --map-only

Speed tip: The first full run extracts all time windows from GEE (slow, ~1 min/window). Every subsequent run loads the cached timeseries_{region}.csv instantly and skips GEE. Use --force-ts to refresh the cache after changing the temporal config.

Outputs

File	Description
`src/timeseries_{region}.csv`	Per-window AOI-mean values for all bands + RRI (cached)
`src/timeseries_rri_{region}.png`	RRI time series
`src/timeseries_ndvi_{region}.png`	NDVI time series
`src/timeseries_ndre_{region}.png`	NDRE time series
`src/timeseries_nbr_{region}.png`	NBR time series
`src/histogram_recovery_class_{region}.png`	Pixel counts by recovery class
`src/map_{region}.html`	Interactive Folium map — open in any browser

HTML map

Toggle any layer → the legend in the bottom-right updates automatically.
Toggle any layer → the time-series panel (top-left) shows the matching index over time with a reference-event line (RRI layer → RRI chart; NDVI layer → NDVI chart; etc.).
Click a red cluster marker (failed-pixel diagnostics) → popup with reasoning and an interactive time-series chart for that location.

Config

All pipeline parameters live in a YAML file under data/. Key sections:

spatial:
  type: geojson
  geojson_path: area_pt.geojson
  region_name: portugal-aoi

temporal:
  start: "2018-01-02"
  end:   "2024-01-01"
  cadence:
    type: fixed
    interval: monthly

comparison:
  prefire_years:  [2017, 2017]   # pre-fire composite window
  prefire_months: [4, 6]
  postfire_years: [2017, 2017]   # fire-scar composite window
  recent_years:   [2022, 2025]   # recent recovery composite
  months:         [5, 9]

classification:
  rri_good_threshold:      0.80  # ≥ this → recovering well
  rri_low_threshold:       0.30  # < this → not recovering
  burn_severity_threshold: 0.10  # dNBR floor to restrict to burned pixels

Architecture

gee_config.py       CLI args, YAML parsing, GEE init, AOI, date windows
gee_s2.py           Sentinel-2 composite builders, spectral indices, biophys proxies
gee_recovery.py     RRI, recovery_class, guard_flags, forest mask
gee_diagnostics.py  Failed-cluster sampling, TS extraction, diagnostic plots
gee_sensitivity.py  Threshold sensitivity analysis
gee_map.py          Folium map builder, dynamic legend, Chart.js TS panel
landsat_lst.py      Landsat 8/9 LST median, anomaly, and stress class
gee.py              Orchestrates all of the above

Spectral indices

Index	Formula (S2 SR)	Interpretation
NDVI	(B8−B4)/(B8+B4)	Vegetation greenness
NDRE	(B8A−B5)/(B8A+B5)	Chlorophyll / canopy quality
NDMI	(B8−B11)/(B8+B11)	Canopy water content
BSI	((B11+B4)−(B8+B2))/((B11+B4)+(B8+B2))	Bare soil exposure
NDWI	(B3−B8)/(B3+B8)	Open water
NBR	(B8−B12)/(B8+B12)	Burn severity / fire recovery

Recovery classes

Class	Color	Meaning
0	Grey	Uncertain — masked, water, or insufficient data
1	Dark green	Recovering well (RRI ≥ 0.80, NDRE confirms canopy)
2	Light green	Recovering but weak
3	Red	Not recovering (RRI < 0.30)
4	Brown	Outside burn recovery assessment (unburned)

Citation / acknowledgements

Sentinel-2 data: ESA Copernicus COPERNICUS/S2_SR_HARMONIZED via Google Earth Engine
Cloud masking: Cloud Score Plus (GOOGLE/CLOUD_SCORE_PLUS/V1/S2_HARMONIZED)
Forest mask: Dynamic World V1 (GOOGLE/DYNAMICWORLD/V1)
LST: Landsat Collection 2 Level-2 (LANDSAT/LC08/C02/T1_L2, LANDSAT/LC09/C02/T1_L2)
RRI methodology adapted from Bright et al. (2019), Remote Sensing of Environment

FilesExpand file tree

README.md

Latest commit

History