Climate change has had a significant impact on natural disasters, particularly floods, in recent times. Early warning systems play an important role in river flood prediction. However, cloudbursts trigger urban flash floods, causing significant disruption to humanity, property damage, and loss of life. In particular, the number of deaths from urban floods has increased in recent times, primarily due to a lack of information. Urban floods, in particular, inflict damage on assets such as vehicles, electric poles, and plants. In addition, flash floods in urban areas submerge roads, drainage, etc., leading to drowning and fatalities. Currently, there is a need to develop smart urban flood prediction and monitoring systems that disseminate instant flood information to rescue teams for a quick response. Currently, deep learning technologies play a significant role in object prediction, but their accuracy in predicting urban flood objects is relatively low. In deep learning algorithms, the training of networks, in conjunction with optimizers and epochs, plays a crucial role in achieving higher accuracies in object detection. The current article investigates the best deep learning training networks, optimizers, and epochs to train urban flood data objects that can achieve higher accuracy. This study considers two pre-trained models, XceptionNet and AlexNet, and three optimizers, including SGDM, ADAM, and RMSProp, to train the urban flood dataset, ensuring balance. We evaluate each training network's performance with the optimizer by tuning epochs and hyper-parameters as constants. Specifically, applying XceptionNet to the SGDM optimizer resulted in an accuracy of 97.47%. The results show that XceptionNet outperforms AlexNet in terms of performance and is recommended for flood object classification. Currently, the study solely focuses on two pre-trained networks and achived 97.47% accuracy; however, it has the potential to evaluate other deep convolutional neural networks, potentially achieving 100% data object training.