Beyond Backscatter: InSAR coherence from detected SAR images

Deep learning framework for estimating InSAR coherence directly from detected SAR imagery, enabling large-scale coherence generation from SAR products available through Google Earth Engine.

This repository will contain the code, experiments, notebooks, pre-trained models, and presentation slides accompanying the paper "Beyond Backscatter: InSAR Coherence from Detected SAR Images" from Francescopaolo Sica, Andrea Pulella, and Michael Schmitt.

Google Colab notebook

The public Colab notebook estimates InSAR coherence from two Sentinel-1 GRD/GEE backscatter images:

https://colab.research.google.com/github/FPSica/BeyondBackscatter/blob/main/notebooks/back2coh_grd_gee_colab.ipynb

The user chooses a region of interest and two acquisition date windows, authenticates Google Earth Engine, selects one exact Sentinel-1 GRD image per period, downloads the shared valid coverage area, and runs the GRD/GEE Beyond Backscatter model workflow. This notebook uses Sentinel-1 GRD data from Google Earth Engine, not SLC inputs.

Model weights are loaded from Hugging Face Hub by default. Google Drive and local model directories are also supported.

Predicted coherence is displayed in grayscale. The RGB product generated by the notebook is a SAR/coherence diagnostic visualization derived from SAR amplitudes and predicted coherence. It is not true optical imagery and not an optical reconstruction.

Model weights

Model checkpoints are not stored directly in this GitHub repository.

Recommended public model repository:

https://huggingface.co/FPSica/beyond-backscatter-grd-gee

The notebook expects the Hugging Face model repository to contain the real GRD/GEE TensorFlow/Keras weights and configuration, by default:

• model.weights.h5
• config.yaml
• README.md

The matching TensorFlow/Keras ResUNet architecture is included in this GitHub repository under src/colab_grd_gee/tf_model.py. The loader first tries standard Keras weight loading and falls back to legacy Keras H5 by-name loading for newer Keras runtimes.

If you mirror or replace the model repository, set HF_REPO_ID in the notebook to the correct released model repository.

🚧 Work in progress: this repository is still evolving, and additional experiments, documentation, and utilities may be added over time.

📑 Presentation slides: available in the slides/ directory of this repository.

📄 Preprint available on Arxiv: https://www.researchgate.net/publication/406351207_Beyond_Backscatter_InSAR_coherence_from_detected_SAR_images