floodsr: A lightweight Python toolkit for terrain-conditioned flood map super-resolution

Coarse flood maps support regional risk assessment but omit terrain detail relevant to local exposure. We present floodsr, a Python framework for reconstructing fine-grid flood depths using a trained residual U-Net (ResUNet) or the existing terrain-based CostGrow method. We trained the terrain-conditioned ResUNet on paired Canadian hydraulic simulations and evaluated it on withheld spatial tiles and independent simulations at two external sites. Wet-loss training improved reconstruction on five selected tiles, although internal validation favoured the deployed baseline. External depth errors decreased, but wet-depth errors and spatial agreement did not improve consistently. CPU benchmarks quantified runtime and memory for both methods, and repeatability and input-validation checks passed in the tested configurations. The framework supports comparison and reuse of reconstruction methods; application-specific validation remains necessary because finer grids do not guarantee more accurate flood maps. Non-peer-reviewed preprint. Includes the manuscript and supplementary material. Software: https://github.com/cefect/floodsr . Documentation: https://floodsr.readthedocs.io/en/latest/ .