From pixels to runoff: integrating high-resolution remote sensing and deep learning for hydrologically meaningful urban land-use classification

Accurate urban land-use classification is essential for hydrological modeling, yet most studies evaluate classification schemes solely by pixel-level accuracy rather than their hydrological consequences. This study examines whether conventional classification accuracy is sufficient to identify land-use maps suitable for hydrological modeling. To address this question, high-resolution (8 cm), 3-band aerial imagery was classified using two machine-learning classifiers, random forest (RF) and support vector machine (SVM), and three deep-learning semantic-segmentation models (U-Net, PSP-Net, DeepLabV3). Classification accuracy was evaluated against reference points using Overall Accuracy (OA), Kappa, and class-specific accuracies. Hydrological performance of the resulting maps was assessed with NSE, R 2 , KGE, and RMSE. U-Net achieved the highest macro-IoU (63.1%) and stronger spatial agreement for Pavement and Sidewalk, indicating better preservation of hydrologically influential impervious features. Although postprocessed RF and U-Net achieved equivalent OA values of 84.7%, DL-based classifications generally produced stronger hydrological performance, with U-Net frequently among the strongest-performing models. Design-storm analysis revealed that the type of classification altered peak flow by 1.54 m 3 /s, demonstrating the sensitivity of runoff predictions to the classified land-use input. These findings demonstrate that overall pixel-level accuracy alone is insufficient to determine hydrological suitability and that class-specific representation of hydrologically important surfaces should also be considered. This study highlights the importance of evaluating remote sensing-based classification with respect to hydrological outcomes, with implications for stormwater infrastructure design and urban flood resilience.