We would like to congratulate lab member Maisha Mahboob on the recent successful completion of her Master's of Science. Read her thesis abstract below:
Dual-Frequency Polarimetric SAR Analysis and Physics-Constrained Neural Networks for Arctic Sea Ice Classification: A NISAR Baseline Study
Satellite based remote sensing has been the primary tool to monitor this dynamic sea ice in the new regime, among them the NASA–ISRO Synthetic Aperture Radar (NISAR) mission provides the first spaceborne dual-frequency (L- and S-band) fully polarimetric SAR capability for cryosphere monitoring. This thesis presents a comprehensive investigation of Arctic sea ice classification using coincident L- and S-band polarimetric SAR data, establishing both the physical foundations and machine learning frameworks for operational ice monitoring in the NISAR era.
The first study characterizes standard polarimetric parameters across dual frequencies for sea ice discrimination using airborne UAVSAR imagery acquired over the Beaufort Sea during winter 2019 and summer 2021. Radar data analysis of new ice (NI), first-year ice (FYI), and multi-year ice (MYI) reveals that total power (SPAN) is the most robust discriminator across ice types, seasons, and frequencies. Advanced radar decomposition technique such as Freeman–Durden decomposition reveals a novel frequency-dependent finding: S-band MYI scattering is surface-dominated (43%), whereas L-band MYI is volume-dominated (44%), demonstrating complementary sensitivities to ice properties. This work also found that only three advanced radar parameters SPAN, mean alpha angle (ᾱ), and co-polarization correlation coefficient (ρ) is enough to achieve the highest accuracy in sea ice classification.
The second study introduces a novel Physics-Constrained Neural Network (PCNN) that eliminates label dependency by embedding established SAR–ice scattering relationships as ten differentiable physics-based loss constraints, thereby enforcing physically consistent class assignments during training without requiring manual annotations. Over a complex sea ice regime in Beaufort sea, PCNN achieves 71.6% overall accuracy, outperforming supervised Random Forest (65.8%) and unsupervised Gaussian Mixture Models (58.9%).
This new model has higher efficiency and accuracy to detect thin ice in refrozen leads. Together, these contributions establish physics-based foundations and scalable, label-free classification frameworks for dual-frequency sea ice monitoring, with direct applicability to the recently launched NISAR mission assuring safe improved sea ice monitoring in the Arctic.
