Completing the machine learning saga in fractional snow cover estimation from MODIS Terra reflectance data: Random forests versus support vector regression

Abstract This study; i) investigates the suitability of two frequently employed machine learning algorithms in remote sensing, namely, random forests (RFs) and support vector regression (SVR) for fractional snow cover (FSC) estimation from MODIS Terra data, and ii) compares them with the previously proposed artificial neural networks (ANNs) and multivariate adaptive regression splines (MARS) methods over an heterogeneous and complex alpine terrain. The dataset comprises 20 Landsat 8 – MODIS image pairs that belong to European Alps acquired from Apr 2013 to Dec 2016. The fifteen image pairs are used to generate the training dataset necessary to build the models, whereas the remaining five are employed as a separate test dataset. The reference FSC maps are derived from the binary classified Landsat 8 snow/no snow maps at 30 m resolution. In order to assess the effect of sampling type and sample size, nine different training datasets are generated. The RF and SVR models are trained accordingly by using various settings of model tuning parameters. During the training of the models, MODIS top-of-atmosphere reflectance values of bands 1–7, NDSI, NDVI and land cover class are input as independent variables (i.e., predictors) to estimate the dependent variable (i.e., response), i.e., FSC value. The resolution of the generated FSC maps is 500 m. The results indicate that the ANN, MARS, RF and SVR models exhibit high consistency with reference FSC values as indicated by low RMSE (~0.14) and high R (~0.93) values. In order to analyze the effect of using three auxiliary variables, i.e., NDSI, NDVI and land cover class, to the predictive ability of the models; ANN, MARS, RF and SVR models are also trained without these predictor variables, i.e., by only using MODIS bands 1–7. The models trained without three auxiliary variables slightly differ from the ones trained with the full set of predictors by only resulting in a mean decrease in R