A snow depletion curve (SDC), the relationship between snow mass (e.g., snow depth [SD]) and fractional snow cover area (SCF), is essential to parameterize the effect of snowpack within a physically based snow model. Existing SDCs are constructed using traditional statistic methods may not be applicable in complex mountainous areas. In this study, we developed an information fusion framework to define the relationship between SCF and SD as well as 12 auxiliary factors by using a traditional statistical method and four prevailing machine learning (ML) algorithms, which have comprehensively considered the variable conditions that cause spatiotemporal heterogeneity of snow cover. We also performed a single-dimensional sensitivity analysis to investigate the physical rationality of the newly developed SDCs. The Northern Xinjiang, Northwest China, is selected as the study area, and the data from 46 meteorological stations covering five snow seasons from 2010 to 2015 are used. The results illustrated that ML techniques can be used to establish high-accuracy and robust SDCs for complex mountainous areas. Compared with SDCs constructed by traditional statistical, the performance of the four ML-based SDCs is significantly improved, the RMSE values can be reduced by 50%, R² above 0.75, and an average relative variance close to 0. ML-based SDCs predicted SCF values showed a range of sensitivities to different input variables (e.g., Land surface temperature, aspect, longwave radiation and land cover type), in addition to SD, that were physically representative of effects that snow cover is sensitive to. Moreover, the complexity of SDCs can be reduced by removing insensitive input variables.