Floods impose severe risks in urban areas, yet operational mapping often stops at inundation extent rather than depth, which is critical for assessing and managing accessibility, and risk. Progress in deep learning offers a path forward but is constrained by the scarcity of large quantity, georeferenced depth datasets that are well labelled. We present Inundation2Depth, a dataset that pairs inundation extent-depth labels derived from aerial imagery and LiDAR-based DTMs (Digital Terrain Models) under hydrostatic assumptions (water-surface elevation relative to terrain). The release includes complementary layers from multi-sensor remote sensing (e.g., DTM-derived terrain features, such as slope, curvature, etc., and land surface characteristics, such as impervious surfaces, vegetation). Data are provided as scene-level raster and 256×256 tiles, in Raw and Normalized versions with consistent naming to support direct use in machine/deep learning pipelines. It constitutes a total of 5,925 overlapping tiles. The dataset enables benchmarking and training of depth-regression models in urban settings, facilitates ablation studies on feature utility, and supports reproducible workflows for hazard assessment. Inundation2Depth lowers the data barrier for GeoAI research on urban flood severity and promotes comparability across methods and study areas. The generated dataset was validated through a hydrodynamic modeling approach using HEC-RAS Rain-on-Grid tool.