Abstract： Accurate retrieval of snow depth from passive microwave observations remains a major challenge because of the pronounced nonlinearity between microwave brightness temperature and snow physical properties, which is further modulated by complex terrain and the temporal evolution of snowpacks. Conventional retrieval approaches, including physically based radiative transfer models, are often constrained by high computational costs and strong sensitivity to uncertain input parameters, resulting in degraded performance under heterogeneous snow conditions and across different snow climate regimes. To overcome these limitations, this study develops a passive microwave snow depth retrieval model that integrates wavelet transform with a residual convolutional neural network (Wavedec-ResNet-CNN) to better represent multi-scale features and complex nonlinear relationships. Specifically, the wavelet transform is applied to multi-frequency brightness temperature observations to extract scale-dependent information, allowing the model to capture both large-scale snow depth variability and localized perturbations associated with rapid snow accumulation and melt processes. The resulting multi-scale features are then ingested into a residual convolutional neural network, in which residual connections facilitate deep feature propagation and enhance nonlinear fitting capability while mitigating performance degradation in deep architectures. The proposed model is evaluated using representative groundbased stations spanning different snow climate classes in the Northern Hemisphere to assess its applicability under diverse snow conditions at the site scale. Model performance is quantified using multiple statistical metrics, including root mean square error (RMSE), correlation coefficient (R), mean bias, and Nash-Sutcliffe efficiency (NSE), and is benchmarked against conventional convolutional neural network and residual convolutional neural network models. The results show that the Wavedec-ResNet-CNN achieves an RMSE of 0.23 m, an R of 0.93, a mean bias of 0.01 m, and an NSE of 0.86, indicating consistent improvements over the reference deep learning models across all evaluation metrics. Additional analyses demonstrate that the proposed model outperforms traditional deep learning approaches across a wide range of snow depths, with particularly notable gains during periods of deep snow accumulation and rapid snow depth changes. Furthermore, compared with existing snow depth retrieval products, the model exhibits enhanced temporal responsiveness and improved spatial transferability across different snow climate types, highlighting its potential for robust passive microwave snow depth monitoring in regions with diverse snow regimes.