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

Abstract

Deep learning segmentation models achieve strong source-domain performance for sand dune mapping, yet their transferability across mineralogically distinct desert environments remains insufficiently characterized. This study evaluates four U-Net architectures — Optical, SAR, Early Fusion, and Attention Fusion — trained on Sentinel-1 and Sentinel-2 imagery from a Moroccan source domain under zero-shot transfer, unsupervised domain adaptation, and few-shot fine-tuning across three target environments: Morocco East, Oman, and Lyndon River West, Australia. Under zero-shot transfer, optical-informed architectures collapsed to near-zero IoU at mineralogically inverted sites (Morocco East: 0.000–0.110; Australia: 0.003–0.029), whilst the SAR U-Net achieved the sole above-baseline result at Morocco East (IoU = 0.625; +36.0%). Scalar distributional distance did not predict transfer degradation (r = −0.168, p > 0.05), whereas the sign of the normalized Sentinel-2 B12 mean was associated with all transfer failure outcomes. Unsupervised polarity inversion recovered optical model performance to the class ratio boundary without labelled data. Five-shot fine-tuning elevated optical-informed architectures above source-domain performance at Morocco East (mean IoU = 0.627–0.729 at N = 5), whilst no architecture exceeded the class ratio at Australia at any shot count, confirming a persistent performance ceiling driven by compound spectral and morphological domain shift. The proposed B12 polarity diagnostic correctly identifies transfer failure risk at a normalized mean threshold of −1.5, providing a zero-label pre-deployment screening tool. Monte Carlo Dropout entropy–error correlations inverted from positive at Oman to negative at Australia; few-shot fine-tuning partially restored calibration at Australia (Spearman ρ = +0.145 to +0.319 at N = 20).