Reproducibility Package for: "Spectral Contrast Polarity Inversion as a Transfer Failure Mechanism in Deep Learning Sand Dune Segmentation: A Cross-Desert Evaluation Using Sentinel-1 and Sentinel-2" (Manuscript RSASE-D-26-00674)

his repository contains the complete dataset and computational pipeline required to fully reproduce the experimental results presented in the manuscript RSASE-D-26-00674. It is designed for strict, end-to-end reproducibility, encompassing everything from raw data preprocessing to the generation of final publication figures.

Contents:

• Data (
  data/): Raw Sentinel-1 and Sentinel-2 raster stacks and ground truth shapefiles for the study sites (Morocco, Oman, Australia), alongside the scale-harmonized Top-of-Atmosphere (TOA) reflectance patches (64×64) ready for model training.
• Pipeline (pipeline/): A sequentially numbered (19-step) Python pipeline that automates:
  1. Multi-modal training and 5-fold Cross-Validation of four U-Net architectures (Optical, SAR, Early Fusion, Attention Fusion).
  2. Global Zero-Shot transfer evaluation.
  3. Spectral domain shift diagnostics (Wasserstein Distances and Threshold Sensitivity).
  4. Unsupervised (Contrast Inversion, Histogram Matching) and Supervised (Few-Shot) Domain Adaptation.
  5. Uncertainty Quantification (MCDropout Calibration).
  6. Automated generation of all publication figures, Grad-CAM activation maps, and performance tables.

Note: In accordance with rigorous reproducibility principles, this package provides only the source data and executable scripts. No pre-computed results are included. Users must execute the pipeline locally to generate all metrics, statistical analyses, and figures from scratch. Detailed execution instructions are available in the included

README.md and 

REPRODUCIBILITY.md files.