Mangrove sub-canopy topography plays a critical role in coastal hydrological processes, blue carbon storage, ecosystem stability, and inundation vulnerability under sea-level rise. However, existing global Digital Elevation Models (DEMs) often contain large elevation uncertainties and data gaps in mangrove regions because dense canopy cover limits the penetration capability of conventional remote sensing observations, resulting in incomplete and inaccurate representations of sub-canopy terrain. To address this critical data deficiency, we present RFDTM, a large-scale mangrove sub-canopy topography dataset for New Zealand at 30 m spatial resolution generated entirely from publicly available satellite observations. The dataset was developed by integrating ICESat-2 photon-counting LiDAR data with dual-frequency C-band and L-band SAR observations. First, a Hierarchical Multi-Constraint Filtering (HMCF) strategy was employed to extract reliable ground photons and improve the reliability of terrain elevation estimates beneath dense canopies. Subsequently, multi-source terrain and vegetation features were constructed and optimized within a Random Forest regression framework to reconstruct continuous sub-canopy topography and generate the RFDTM product. Validation against airborne LiDAR terrain data across all mangrove regions of New Zealand demonstrates excellent performance, with an R² of 0.99, RMSE of 1.01 m, MAE of 0.80 m, and bias of 0.43 m, fully satisfying the accuracy requirements for regional-scale applications. Ablation experiments further confirm the critical contribution of L-band SAR observations, reducing the RMSE from 1.23 m to 1.01 m and substantially enhancing sub-canopy penetration capability. Overall, RFDTM represents the first large-scale mangrove sub-canopy topography product derived solely from open-access satellite data, while the proposed methodology provides a transferable and readily applicable framework for global coastal vulnerability assessment, ecosystem monitoring, and carbon cycle studies.