Estimating snow depth by combining satellite data and ground-based observations over Alaska: A deep learning approach

Abstract

Snow cover plays a vital role in the climate system because it is related to climate, hydrological cycle, and ecosystem. On this basis, deriving a long-term and large-scale snow depth (SD) time series and monitoring its temporal and spatial variations are crucial. Passive microwave remote sensing data in combination with in-situ SD data have long been used to retrieve SD. However, the retrieval accuracy is limited in case of sparse meteorological stations, and the high-quality applications of retrieval results are hindered in specific areas. The ground-based global navigation satellite system reflectometry (GNSS-R) method is currently a potential way to monitor SD variations with a high degree of accuracy but has a limited spatial coverage. In this study, a deep learning-based approach, which displays a stronger nonlinear expressiveness capability than conventional neural networks, was applied to estimate SD by combining satellite observations, in-situ data, and GNSS-R estimates. The model was trained and tested with data obtained in Alaska between 2008 and 2017. Results show that the proposed deep belief network model performs better than linear methods and conventional neural network models and demonstrate the effectiveness of combining GNSS-R estimation with increased cross-validation R of 0.85 and decreased RMSE of 15.40 cm. The predicted SD distribution indicates that the variations in mean SD in Alaska for March and April between 2008 and 2017 were associated with the climate anomalies and air temperature. Overall, the proposed deep learning-based method is a promising approach in the satellite-retrieved SD field.