Comparative Analysis of DeepLabv3+ for Urban Waterbody Mapping: Enhancing Policy-Making and Sustainable Urban Development

Waterbody extraction plays a crucial role in various applications such as environmental monitoring, urban planning, and disaster management. Over the past decade, numerous convolutional neural networks (CNNs) have been developed specifically for the accurate segmentation of waterbodies. However segmenting tiny branch-like structures of waterbodies observed in images remains challenging. DeepLabV3 + is indeed one of the top segmentation models excelling in the task of segmenting tiny waterbody structure. However, its computational demands are a major drawback. Therefore, this paper investigates the performance of deepLabV3 + using various backbone networks such as EfficientNet, MobileNet, ResNet50, DenseNet121, and YOLOv8. Among the selected backbone networks, EfficientNet achieves excellent accuracy with relatively efficient computation time because of its compound scaling approach. It surpasses DenseNet by 1.2%, outperforms ResNet50 by 1.62%, achieves 1.86% higher accuracy than MobileNet, and significantly exceeds YOLOv8 by 3.71%. Experimental results demonstrate that deepLabV3 + using EfficientNet stands out as the most effective segmentation model, achieving the highest Structural Similarity Index (SSIM) value of 0.963 and lowest mean absolute distance (MAD) value of 0.891, particularly for customized data segmentation of water bodies with tiny branch-like patterns.