Artificial Neural Network Applications for Investigating Microwave Radiative Transfer in Polar Environments

Abstract

Accurate retrievals of surface and atmospheric parameters in polar regions are essential for climate monitoring, yet these environments pose unique challenges for remote sensing. Snow and ice exhibit strong spatial and temporal variability in properties such as depth, density, and grain size, leading to large variations in microwave penetration depth which complicates the interpretation of passive microwave satellite observations. This variability, combined with limited in situ observations, makes polar regions particularly difficult to characterize using traditional retrieval methods. This thesis explores whether Artificial Neural Networks (ANNs) can improve retrievals of geophysical parameters from passive microwave observations made by the Advanced Technology Microwave Sounder (ATMS), while also using these retrievals to investigate the physical properties of snow and ice.

The study develops and validates two ANN-based retrieval approaches: a single-pixel Multi-Layer Perceptron (MLP) model and an image-to-image Swin-UNet model, which incorporates spatial context. Both models are trained using RTTOV-simulated brightness temperatures and validated against independent ground-based datasets, including measurements made by radiosondes, pyrgeometers, and ice mass balance buoys. The single-pixel retrievals demonstrate strong agreement with in situ observations for atmospheric profiles, emissivities, and frequency-dependent effective surface temperatures, while the image-to-image model further improves spatial coherence and retrieval stability by exploiting spatial structure in satellite swaths. Comparisons with ERA5 show that these ANN approaches often achieve comparable performance using this single-instrument retrieval, while also providing additional surface information in the effective surface temperatures and effective emissivities at lower frequencies.

Finally, the retrieved effective surface temperatures and emissivities are compared to physics-based radiative transfer simulations using the Improved Born Approximation (IBA) and the Snow Microwave Radiative Transfer (SMRT) model. These comparisons demonstrate that the ANN retrievals encode physically meaningful information about snow. In particular, the 23.8 GHz effective surface temperature can be used as a proxy for the snow-ice interface temperature for snow-covered ice, while the 183.3 GHz effective surface temperature provides a reasonable approximation for the snow-air interface temperature, thereby providing estimates of the snow temperature gradients. Additionally, the retrieved emissivities can be used to interpret surface and snowpack conditions, where their frequency-dependent behaviour reflects physical changes in the surface state, providing qualitative indicators of evolving snow and ice surface conditions. By capturing the physical relationships between microwave emissions and snow properties, these ANN-based retrievals open new avenues for monitoring polar environments and refining our understanding of their role in the Earth’s climate system.