siyux1927/imaestro-sar-forest

I-MAESTRO SAR Forest Patches

Overview

This dataset contains processed Sentinel-1 SAR imagery and forest structural attributes derived from the I-MAESTRO synthetic forest inventory, covering three European sites. It was prepared for multi-task learning experiments on joint forest genus segmentation, canopy height regression, and aboveground biomass regression.

Source

Forest inventory: I-MAESTRO synthetic dataset — tree-level inventory data for three European forest landscapes (Bauges/France, Milicz/Poland, Sneznik/Slovenia), covering ~100,000 ha with 42M+ trees from 51 species at 25m grid resolution.

SAR imagery: Sentinel-1 GRD (Ground Range Detected) dual-polarization scenes acquired during the 2019 growing season, downloaded from the Alaska Satellite Facility (ASF). Processed with ESA SNAP: radiometric calibration, terrain correction (SRTM DEM), reprojection to local CRS, and resampling to 25m.

Biomass targets were estimated from tree-level data using species-specific allometric equations. Canopy height uses the 95th percentile height per cell. Dominant genus was determined by tree count per cell.

Dataset Structure

training-patches/
  patches_{train|val|test}.npy          # (N, 5, 64, 64) float32
  labels_dominant_genus_{split}.npy     # (N,) int — dominant genus code
  sites_{split}.npy                     # (N,) str — site name

rasters/
  {bauges|milicz|sneznik}/
    {site}_vh.tif                       # VH backscatter (dB)
    {site}_vv.tif                       # VV backscatter (dB)
    {site}_biomass_t_ha_smooth.tif      # Biomass (t/ha)
    {site}_height95_smooth.tif          # Canopy height 95th pct (m)
    {site}_dom_genus_smooth.tif         # Dominant genus (integer code)

auxiliary/
  genus_map.json                        # genus name → integer code
  wood_density_allometric.json          # allometric parameters per species
  split-blocks/
    {train|val|test}_{site}_blocks.geojson  # spatial block boundaries

Patch Format

Each .npy patch array has shape (N, 5, 64, 64) with channels in fixed order:

Channel	Variable	Unit
0	VH backscatter	dB
1	VV backscatter	dB
2	Biomass (smoothed)	t/ha
3	Dominant genus (smoothed)	integer code
4	Canopy height 95th pct (smoothed)	m

Channels 0-1 are model inputs; channels 2-4 are prediction targets. Invalid pixels are NaN.

Splits: Train 724 patches / Val 145 / Test 109. Patches were extracted with a 50% sliding window overlap within spatially blocked regions (4×4 patch blocks, ~3.2×3.2 km each) to prevent spatial autocorrelation between splits.

Usage

import numpy as np

patches = np.load("training-patches/patches_train.npy")   # (724, 5, 64, 64)
labels  = np.load("training-patches/labels_dominant_genus_train.npy")
sites   = np.load("training-patches/sites_train.npy")

# SAR inputs
X = patches[:, :2, :, :]   # VH, VV

# Regression targets
biomass = patches[:, 2, :, :]
genus   = patches[:, 3, :, :]
height  = patches[:, 4, :, :]

Genus integer codes are defined in auxiliary/genus_map.json. The six dominant genera used in training are: Abies (1), Fagus (11), Fraxinus (13), Picea (19), Pinus (20), Quercus (25).

Sites

Site	Country	CRS	Forest area	Dominant genera
Bauges	France	EPSG:2154	~45,000 ha	Fagus, Abies, Pinus
Milicz	Poland	EPSG:2180	~20,000 ha	Pinus, Quercus
Sneznik	Slovenia	EPSG:3912	~5,300 ha	Fagus, Abies

License

Dataset released under CC-BY 4.0. SAR data sourced from Copernicus Sentinel-1 (ESA, open access). I-MAESTRO inventory data used for research purposes.