Natural calamities, including floods, can cause severe damage to the environment, human life, and property. Accurate and timely identification of inundation areas is essential for effective and efficient disaster preparedness, response, and recovery efforts. In recent years, deep learning algorithms, such as Convolutional Neural Network (CNN), have achieved tremendous success for image-related tasks. Consequently, this paper aims to propose a model based on CNN for early detection of flood disasters through image segmentation. The proposed solution leverages the potential of UNet architecture and employs EfficientNetB7 as a backbone to improve the accuracy in detecting flood-affected areas. Using the encoder-decoder architecture, the proposed approach takes an image as an input and generates its corresponding segmentation mask. The model was trained and tested on the HISEA-1 Synthetic Aperture Radar (SAR) dataset using a hold-out validation method and benchmarked with various baseline architectures using several evaluation metrics. The experimental results showed that the proposed solution outperformed the existing approaches in the flood detection domain, achieving the highest Accuracy, F1-Score, and IoU values of 95.99%, 97.53%, and 95.19%, respectively. The proposed model also improved the detection accuracy of the state-of-the-art solutions by 1.99% to 3.99%. Through accurate flood image segmentation, this study possesses the potential to facilitate the enhancement of disaster management systems.