Abstract: Snow is an important component of the cryosphere, affecting the energy balance and water exchange between land, ocean, and atmosphere. This article discusses the feasibility of snow density inversion based on Global Navigation Satellite System Reflectrometry (GNSS-R) satellite TechDemoSat-1 (TDS-1) delay-Doppler map (DDM) data from 2016 to 2018. Select different proportions to divide the experimental data from 2016 and 2017 into training and validation sets, use the training set to establish machine learning algorithms such as random forest (RF) and extreme gradient enhancement, as well as convolutional neural network deep learning algorithms for snow density inversion models. Evaluate the model performance through the validation set The experimental results show that RF has the best inversion performance, and performs the best when dividing the data in a 7∶3 ratio. Its mean absolute error (MAE) is 18.021 5 kg/m³, root mean square error (RMSE) is 28.700 4 kg/m³, and the determination coefficient R² is 0.818 5. On this basis, the generalization performance of the established RF model was evaluated using 2018 data as the test set. The MAE was 42.690 9 kg/m³, RMSE was 54.438 0 kg/m³, and R² was 0.285 9. The feature parameters are divided into three categories. The results of Shapley additive explanations (SHAP) analysis and ablation experiments show that TDS-1 features play a major role in the inversion model, while environmental features such as snow layer temperature and wind speed play important auxiliary roles. TDS-1 derived features further improve the accuracy of the model. Analysis shows that the distribution of data directly affects the performance of the model, and the model has the best inversion accuracy in the snow density range of 150–250 kg/m³, where the data proportion is the highest. The experimental results have demonstrated the feasibility of using TDS-1 satellite data to invert snow density, providing new ideas for snow density inversion research.