Beyond Backscatter GRD/GEE Coherence Estimator
This model package contains the TensorFlow/Keras GRD/GEE model weights for Beyond Backscatter: InSAR Coherence from Detected SAR Images.
Task
Predict an InSAR-like coherence map from two detected Sentinel-1 GRD/GEE SAR backscatter images.
Inputs
The public notebook downloads two Sentinel-1 GRD sigma0 backscatter images from Google Earth Engine in linear scale. The default polarization is VV.
Earth Engine preprocessing:
- collection:
COPERNICUS/S1_GRD;
- acquisition mode:
IW;
- orbit pass filtering, default
ASCENDING;
- optional relative orbit filtering;
- two user-selected date windows;
- median composite for each date window;
- dB-to-linear conversion using
10 ** (db / 10);
- selected polarization;
- clipped region of interest;
- 10 m output scale by default.
Model preprocessing:
- convert downloaded linear sigma0 back to dB with
10 * log10(linear + eps);
- clip to
[-20, 0] dB;
- normalize to
[0, 1];
- channel order:
[t1, t2];
- tiled inference with 128 x 128 patches and Kaiser-window aggregation.
Outputs
The model outputs a predicted coherence map in [0, 1]. The public notebook saves the map as GeoTIFF, PNG, and NumPy products, preserving georeferencing from the downloaded Sentinel-1 inputs.
Files
model.weights.h5: real GRD/GEE TensorFlow/Keras weights in legacy Keras H5 format.config.yaml: model, preprocessing, tiling, and output conventions.model_metadata.json: lightweight public packaging metadata.
The TensorFlow/Keras architecture implementation is provided by the GitHub repository in src/colab_grd_gee/tf_model.py.
Limitations
- This is not the SLC-based workflow.
- This is not true interferometric processing from complex SLC data.
- The pseudo-RGB products produced by the notebook are SAR/coherence visualizations, not optical imagery.
- Earth Engine authentication and a valid Earth Engine-enabled Google Cloud project are required to run the full notebook.
- Start with a small ROI before processing larger areas.
Citation
Beyond Backscatter: InSAR Coherence from Detected SAR Images
Francescopaolo Sica, Andrea Pulella, Michael Schmitt
License
The GitHub code repository is MIT licensed. The model-weight license should be confirmed by the authors before redistribution or downstream release.
