Water Region Extraction in Aerial Images Using Deep Learning

Water bodies are essential component of Earth’s surface and accurate water region mapping is crucial for disaster management and environmental monitoring. Existing methods to map water bodies relies on traditional modelling methods that need extensive manual calibration struggle with scalability and lack advanced pattern recognition. This paper evaluates the effectiveness of deep learning-based semantic segmentation models, DeepLabV3+ and U-Net, using aerial images resized to $\mathbf{2 5 6 x} 256$ pixels in mapping water regions in aerial images. DeepLabV3+, with a ResNet50 backbone, and U-Net are trained on a comprehensive flood-affected dataset with preprocessing techniques such as resizing and augmentation. Model performance is assessed using metrics like IoU, F1 score, accuracy, precision, and recall. The results indicate that DeepLabV3+ outperforms U-Net, showing promise for automating water region mapping and improving flood surveys and urban planning.