<p>Flood hazard mapping is crucial for managing the impacts of floods, which are among the most frequent and severe natural disasters. Deep learning-based super-resolution techniques offer a fast solution to enhance low-resolution flood maps but face challenges due to dataset imbalance, where dry pixels vastly outnumber flooded ones, leading to overconfident and erroneous predictions. This study investigates the application of entropy regularization to mitigate Convolutional Neural Network’s (CNN) overconfidence in flood map super-resolution. The CNN model was trained using simulated flood depths from a combined fluvial and pluvial flood process. The CNN’s loss function was subtracted by a weighted entropy term, encouraging the model to maintain appropriate uncertainty, especially for the underrepresented high water depth pixels. The effects of this technique on a CNN model's accuracy for flood map super-resolution were examined at factors of × 2, × 4, × 8, and × 16. Our results suggest that higher super-resolution factors, which correspond to more challenging super-resolution tasks, are correlated with the effectiveness of entropy regularization in CNNs. While minimal impact was observed at lower super-resolution factors, entropy regularization improved CNN accuracy at higher factors, suggesting the potential for dynamically adjusting the entropy regularization coefficient based on the super-resolution factor to optimize flood map accuracy. Moreover, similar improvements have been observed at × 8 and × 16 super-resolution factors across other deep learning architectures, including Super-Resolution Generative Adversarial Network (SRGAN), Enhanced Super-Resolution Generative Adversarial Network (ESRGAN), and Multi-Scale Deep Super-Resolution (MDSR).</p>