Preparing S1 + AEF + Dynamic World data

This repository intentionally does not distribute imagery, AlphaEarth embeddings, or labels. Use this guide to create compatible GeoTIFF triplets.

Required rasters per training tile

Raster	Bands	Data type	Meaning
S1 input	3	float32 preferred	`VV`, `VH`, `angle`, in that order
AEF input	64	raw embedding values	Annual AlphaEarth Foundation embedding bands, in source order
DW target	1	integer or float	Binary mask: water=`1`, other=`0`

The training code reads all values as float32, converts non-finite values to zero, and binarizes labels at > 0.5. Therefore, encode water as one and non-water as zero. Do not encode Dynamic World class IDs directly as the target: class ID zero means water, but would be interpreted as non-water by the trainer.

Alignment checklist

For each S1/AEF/label triplet, ensure:

exactly the same CRS, affine transform, width, and height;
exactly one pixel grid, resolution, and area of interest;
S1 is exactly three bands and AEF is exactly 64 bands;
no duplicate filename exists anywhere under the AEF directory;
file names match exactly when using directory-based loading.

The command below is a quick manual inspection for one file. Repeat it for its S1, AEF, and label partners and compare the reported grid fields.

gdalinfo data/s1/tile_0001.tif
gdalinfo data/aef/tile_0001.tif
gdalinfo data/labels/tile_0001.tif

Example directory layout

data/
  s1/
    tile_0001.tif
    tile_0002.tif
  aef/
    tile_0001.tif
    tile_0002.tif
  labels/
    tile_0001.tif
    tile_0002.tif

Or create a manifest from training_manifest.csv and train only a selected sample_id:

python src/train.py \
  --triplet_manifest examples/training_manifest.csv \
  --triplet_sample_id demo_region \
  --artifact_dir runs/demo_region \
  --n_s1_bands 3 --n_aef_bands 64 --n_proj_bands 16 \
  --loss ce_dice --batch_size 4 --crop_size 512

Preparing inference data

Inference needs no Dynamic World target. Provide a directory of three-band S1 GeoTIFFs named s1_YYYY-MM-DD.tif and one 64-band annual AEF GeoTIFF. The inference code reprojects AEF windows to each S1 scene grid, but the S1 files in a single run must share the same CRS, transform, width, and height.

data/inference/
  s1/
    s1_2025-01-15.tif
    s1_2025-01-27.tif
  alphaearth_2025.tif