Snow Depth Retrieval Using Detrended SNR From GNSS-R With Bidirectional GRU

Abstract

Snow depth monitoring is crucial for hydrology, climate research, and avalanche prediction. While traditional global navigation satellite system (GNSS) reflectometer methods offer cost-effective snow thickness retrieval, they suffer from poor accuracy and robustness, especially in complex terrains and extreme weather. This study proposes an innovative snow depth retrieval technique employing a time-series recurrent neural network with bidirectional gated recurrent units (Bi-GRUs). Unlike traditional methods using signal-to-noise ratio (SNR) features, our algorithm utilizes the detrended SNR as Bi-GRU input, aiming to enhance accuracy, particularly in low snow depths and complex terrains. SNR observations from GPS L1 carriers at stations P351 and AB33 were analyzed. The Bi-GRU algorithm demonstrated high consistency with true snow depths at station P351 (coefficient of determination: 0.9766), with the root-mean-square error (RMSE) and the mean absolute error (MAE) of 9.1559 and 6.4185 cm, respectively. Compared to traditional methods, the Bi-GRU model improved the RMSE by 30.9% and the MAE by 44.5%. At station AB33, where snow depth variations were significant, accuracy improvements of 65.6% (RMSE: 7.4905 cm) and 63.2% (MAE: 5.6074 cm) were observed. In addition, the Bi-GRU model exhibited greater robustness compared to long short-term memory. These findings highlight the efficacy of the Bi-GRU-based approach, suggesting its superiority and broader applicability.