A novel approach to retrieving the surface soil freeze/thaw state in the Qinghai-Tibetan Plateau using the seasonality of CYGNSS time series

Soil freeze-thaw (F/T) processes are a typical physical phenomenon on the Qinghai-Tibetan Plateau (QTP), significantly impacting regional climate change and the hydrological cycle. This study presents a Seasonal-Trend Decomposition using Loess and Long Short-Term Memory (STL-LSTM) method to detect spatiotemporal variations in soil F/T on the QTP using time series data from the Cyclone Global Navigation Satellite System (CYGNSS). The model was validated against ERA5 soil temperature data (0-7 cm) and independent in-situ observations, demonstrating good consistency. The SHapley Additive exPlanations (SHAP) model was integrated into the STL-LSTM framework to quantitatively evaluate the contributions of input features to F/T retrieval, revealing that time features contributes the most to retrieval results, followed by surface reflectivity. Moreover, spatiotemporal analysis of QTP F/T dynamics shows prominent seasonal patterns, with topography-induced shielding delaying thawing in central QTP regions and freezing trends extending from low (28 degrees N) to high latitudes (36 degrees N). The proposed method offers a new pathway for monitoring freeze-thaw transitions in high-latitude regions and holds potential for expansion into future high-frequency and multi-polarization GNSS-R missions.