Publication

Abstract : Accurate road extraction from high-resolution satellite imagery is crucial for applications such as urban planning, disaster response, and autonomous navigation. While deep learning-based semantic segmentation models have shown promise, each model tends to specialize in particular feature contexts, leading to variable performance across heterogeneous landscapes. This paper presents MoE-RoadNet, a Mixture of Experts (MoE) approach that combines the strengths of multiple pre-trained segmentation models through a trainable metamodel. Extensive evaluations across three benchmark datasetsMassachusetts road dataset MIT, DeepGlobe (DG), and CHN6CUG (CC)-demonstrate the superior performance of our approach. Specifically, the MoE model achieves an absolute IoU improvement of 1.16% on MIT, 2.91% on DG, and 2.54% on CC. These gains are accompanied by consistent improvements in F1-score, precision, and recall, as well as enhanced continuity and clarity in road predictions.