Deep learning on segmenting large and narrow rivers with aerial RGB imagery: A comparison of convolutional and Vision-Transformer networks

Accurate mapping of water resources from high-resolution RGB imagery is essential for environmental management but presents significant challenges, especially for narrow rivers with variable shapes and spectral complexities. While deep learning methods have advanced this field, a comprehensive performance analysis across different architectural paradigms for this specific task is needed. This study presents an investigation into segmenting rivers of varying widths. First, we evaluate the efficacy of established Convolutional Neural Network (CNN) architectures to benchmark their capabilities in mapping both large (¿10 m) and narrow (¡10 m) rivers. Second, addressing the limitations identified in the initial phase, we assess the state-of-the-art SegFormer (Vision Transformer - ViT) in a direct comparison with the versatile DeepLabV3+ and the Segment Anything Model (SAM). Our findings confirm that while CNNs excel in mapping large rivers, with U-Net achieving a 97.43% F1-Score, their performance deteriorates on narrow watercourses. The subsequent analysis reveals that SegFormer substantially outperforms the CNN benchmark on narrow rivers, achieving an F1-Score of 85.99% and an Intersection over Union (IoU) of 75.42%, representing a crucial 6.86% improvement in IoU over DeepLabV3+. Furthermore, while the Foundation Model (SAM) demonstrated superior sensitivity with a Recall of 91.53%, it suffered from severe over-segmentation, resulting in a remarkably lower IoU of 65.98%. This study demonstrates that while CNNs are effective for well-defined rivers and Foundation Models offer high sensitivity, specialized ViT architectures provide the most robust and precise solution for the automated mapping of varied and narrow river systems. • CNNs excel at mapping large rivers but fail to segment narrow rivers from RGB aerial imagery accurately. • A Vision Transformer-based model (SegFormer) significantly outperforms a highly optimized CNN benchmark (DeepLabV3+). • SegFormer improves the Intersection over Union (IoU) by a critical 6.86% on the challenging narrow river subset. • Vision Transformers offer a robust and cost-effective methodology for creating detailed, large-scale hydrographic maps.