A comparison of machine learning methods for estimation of snow density using satellite images.

Low snow density causes snow to melt quickly, so there is no runoff during the warmer months of the year. Therefore, knowing the snow density can be useful in determining the amount of water. To predict snow density, this study used seven machine learning methods, including adaptive neural‐fuzzy inference system (ANFIS), M5P, multivariate adaptive regression spline (MARS), random forest (RF), support vector regression (SVR), gene expression programming (GEP) and eXtreme gradient boosting (XGBoost). Nine factors expected to affect snow density were considered. These factors were extracted using Google Earth Engine (GEE) from 1983 to 2022. The results showed that the surface temperature had the highest correlation (coefficient = −0.7), and the wind speed had the lowest correlation (coefficient = 0.3) among the considered factors on the snow density. Also, the best method was XGBoost (Nash–Sutcliffe efficiency [NSE] = 0.978, R = 0.957), and the worst method is SVR (NSE = 0.7, R = 0.9). Therefore, snow density can be estimated with good accuracy using a combination of machine learning methods and remote sensing. Highlights: This study used seven machine learning methods to estimate snow density.The surface temperature had the highest influence (R = −0.77) on the snow density.The best method for predicting snow density was XGBoost (NSE = 0.978).The worst method for predicting snow density was SVR (NSE = 0.71). [ABSTRACT FROM AUTHOR]