<![CDATA[Flooding is a natural disaster that has become increasingly frequent and causes significant impacts on urban environments, highlighting the need for rapid and accurate mapping of affected areas. Deep learning–based image segmentation, particularly using the U-Net architecture, has been widely applied for this purpose. However, model performance is not determined solely by network architecture, but is also strongly influenced by the selection of training hyperparameters. This study aims to optimize the performance of a U-Net model with a ResNet50 backbone for flood image segmentation using a Sequential Hyperparameter Tuning approach based on a one-factor-at-a-time scheme. The dataset consists of approximately 3,400 RGB flood images with corresponding binary ground truth masks at an original resolution of 512 × 512 pixels, which are resized to 256 × 256 pixels and preprocessed using CLAHE, gamma correction, and unsharp masking to enhance contrast and boundary clarity of inundated areas. The optimization focuses on optimizer selection, batch size, learning rate, and number of training epochs, as these parameters directly affect convergence stability and segmentation accuracy. Hyperparameter tuning is performed sequentially by evaluating model performance on the validation set using Intersection over Union (IoU) and Dice Similarity Coefficient. Based on this process, the optimal configuration employs the AdamW optimizer, a batch size of 8, a learning rate of 0.00015, and 100 training epochs. Final evaluation is conducted on the test set through retraining with three different random seeds, and performance is reported using mean values. The optimized model achieves a mean IoU of 0.7664 and a mean Dice score of 0.8499, with low standard deviation, indicating stable performance and good generalization capability. These findings demonstrate that systematic hyperparameter optimization plays a crucial role in improving the performance of U-Net ResNet50 for flood image segmentation and provides practical insights for remote sensing–based flood mapping systems.]]>
