GlobalFlow-former: a transformer network based on global context for accurate mainstream line identification in complex river networks

The river mainstream line is a critical parameter for hydrodynamic modeling, ecological environment assessment, and navigation channel management. Its accurate and automatic identification holds significant theoretical and practical importance. However, in complex river networks characterized by bifurcations, mid-channel bars, and wide, shallow sections, traditional image processing methods and Convolutional Neural Network (CNN)-based models often suffer from issues like line fragmentation and misidentification due to the inherent limitations of their local receptive fields. To address this challenge, this paper proposes a Transformer-based network that leverages global context, named GlobalFlow-Former, designed for end-to-end, accurate identification of mainstream lines in complex river networks. We innovatively reformulate this task as an image-to-sequence conditional generation problem. The model adopts an encoder-decoder architecture, utilizing a Swin Transformer to capture global visual features and an autoregressive decoder to generate the coordinate sequence of the mainstream line point-by-point. To validate the model's effectiveness, we propose a strategy for training on large-scale, publicly available geospatial data and evaluate it on an independent, high-quality, manually annotated test set. Experimental results demonstrate the superior performance of our proposed model. On our self-constructed River-Gold-100 test set, its Topological Consistency Score (TCS) reached 0.97, significantly surpassing the 0.85 achieved by advanced segmentation models like Swin-Unet. Furthermore, it reduced the error in Fréchet distance by over 70%, showcasing exceptional robustness and precision.