Publication

Abstract : This study focuses on the problem of accurately delineating connected road structures from high-resolution remote sensing imagery-an important task with broad implications for smart city development, routing systems, and emergency management. Existing convolutional and transformer-based segmentation methods often struggle to capture fine structural details, maintain road connectivity, and preserve directional continuity. In this work, we propose fractional Kolmogorov–Arnold networks (fKANs)-UNet, a novel encoder–decoder architecture designed to address these challenges. The network primarily utilizes directional strip convolutions and a feature selective fusion (FSF) block enhanced by a squeeze-and-excitation (SE) mechanism to refine feature representation. To further improve nonlinear modeling and spectral selectivity, we incorporate fractional Jacobi neural blocks (fJNBs) into the architecture. These blocks perform spectral transformations based on Jacobi’s polynomials of fractional order, enabling the model to effectively learn intricate spatial relationships and structures. To optimize the training, a hybrid objective is utilized, integrating binary cross-entropy (BCE), dice, and boundary-based terms, which collectively enhance pixelwise accuracy and edge consistency. A comprehensive evaluation, including detailed ablation analysis, was carried out using the MIT and DeepGlobe benchmark datasets. Compared to MSMDFFNet, fKAN-UNet achieves a 1.99% gain in IoU and a 1.54% boost in F1 score on the Massachusetts dataset. On the DeepGlobe dataset, it shows a 0.53% increase in IoU along with a 0.35% enhancement in the F1 metric. The code is available at: https://github.com/Jayku88/fKANUNet