Estimating the landscape and soil freeze-thaw (FT) dynamics in the Northern Hemisphere (NH) is crucial for understanding permafrost response to global warming and changes in regional and global carbon budgets. A new framework for surface FT-cycle retrievals using L-band microwave radiometry based on a deep convolutional autoencoder neural network is presented. This framework defines the landscape FT-cycle retrieval as a time-series anomaly detection problem, considering the frozen states as normal and the thawed states as anomalies. The autoencoder retrieves the FT-cycle probabilistically through supervised reconstruction of the brightness temperature (TB) time series using a contrastive loss function that minimizes (maximizes) the reconstruction error for the peak winter (summer). Using the data provided by the Soil Moisture Active Passive (SMAP) satellite, it is demonstrated that the framework learns to isolate the landscape FT states over different land surface types with varying complexities related to the radiometric characteristics of snow cover, lake-ice phenology, and vegetation canopy. The consistency of the retrievals is assessed over Alaska using in situ observations, demonstrating an 11% improvement in accuracy and reduced uncertainties compared to traditional methods that rely on thresholding the normalized polarization ratio (NPR).

This paper presents a convolutional autoencoder deep learning framework for probabilistic characterization of the ground freeze-thaw (FT) dynamics in the Northern Hemisphere to enhance our understanding of permafrost response to global warming and shifts in the high-latitude carbon cycle, using Soil Moisture Active Passive (SMAP) satellite brightness temperatures (TB) observations. The autoencoder recasts the FT-cycle retrieval as an anomaly detection problem in which the peak winter (summer) represents the normal (anomaly) segments of the TB time series. The results demonstrate that the new framework outperforms the widely used fixed-thresholding of the Normalized Polarization Ratio (NPR) by learning the land surface structural and radiometric complexities that might arise in TB times series due to snow cover and vegetation. Validation against ground-based measurements over Alaska shows that the accuracy of the FT-cycle retrievals can be improved by 12%, primarily due to a marked reduction in false detection of short snowmelt episodes as ground thawing by the NPR thresholding approach.