Accurate snow depth (SD) information at global scales is highly important for validation and calibration of hydrological model, snowmelt runoff prediction, and global water supplies estimation. Passive microwave remote sensing has been proven to be an efficient method to detect SD at regional or global scales. However, numerous heterogeneous snow property parameters and forest parameters affect the microwave radiation characteristics, which is a great challenge for SD retrieval. In this study, a novel SD retrieval algorithm that comprehensively considered the multiple sensitive factors was developed at global scales by coupling the Microwave Emission Model of Layered Snowpacks (MEMLS) model and machine learning (ML) technique. The exponential correlation length (Pec), snow density (ρ), and forest cover fraction (

$f_{forest}$

) are three sensitive parameters for retrieving SD from passive microwave observations based on the extended Fourier amplitude sensitivity test (EFAST) algorithm. The key parameter Pec, which characterizes the snow microstructure information inside snowpacks, was optimized by minimizing the brightness temperature between satellite observations and microwave radiative transfer model simulations by integrating the MEMLS, forest impact, and atmospheric radiation attenuation. The random forest (RF) ML technique was used, and the optimal combination of independent variables was selected by correlation heatmap and importance ranking of variables. Six selected essential parameters, including Pec, vertical polarized brightness temperature differences between 18.7 and 36.5 GHz, ρ from ERA5,

$f_{forest}$

, elevation, and longitude together with target SD output were used to train the RF model. Seventy percent of the entire data at meteorological stations were randomly chosen to train the RF model, and the remaining 30% were adopted as the validation data from 2013 to 2020 to ensure the mutual independence of training data and validation data. Compared with independent in situ measurements, the overall RMSE of SD is 13.6 cm, which was noticeably lower than those of the international SD remote sensing products and reanalysis datasets, including AMSR2 (36.2 cm), GlobSnow 3.0 (25.3 cm), ERA5 (38.2 cm), and ERA5-land (29.9 cm). The results demonstrated that the new proposed algorithm, by coupling the MEMLS model with the ML technique, could significantly enhance the overall performance in SD retrieval at global scales.