Machine learning-based snow depth retrieval using GNSS signal-to-noise ratio data

GNSS-IR enables the extraction of environmental parameters such as snow depth by analyzing signal-to-noise ratio, indicating the strength of the GNSS signal. We propose a machine learning (ML) classifcation approach for snow depth retrieval using the GNSS-IR technique. ML classifer algorithms were studied to classify the strong and weak ground refections using input parameters (azimuth angle, satellite elevation angle, day of year, amplitude of refected signal, epoch number, etc.) as independent variables. GPS data collected by UNAVCO AB39 and daily snow depth data from SNOTEL Fort Yukon for a 6-year period (2015–2020) were considered. The frst 4-year data were trained by some well-known ML classifers to weight the input data and then used to classify the strong and weak signals. Tree-based classifers, Random Forest, AdaBoost, and Gradient Boosting overperformed the other classifers since they have more than 70% accuracy, so we performed our analysis with these three methods. The last 2-year data were used to validate both trained models and snow depth retrievals. The results show that ML classifer algorithms perform better results than traditional GNSS-IR snow depth retrieval; they improve the correlations by up to 19%. Moreover, the root-mean-square errors decrease from 15.4 to 4.5 cm. This study has a novel approach to the use of ML techniques in GNSS-IR signal classifcation, and the proposed methods provide a critical improvement in accuracy compared to the traditional metho