Abstract: Imaging systems often face challenges in simultaneously considering both spatial and spectral information. However,existing optical image road extraction methods primarily rely on fused images as data sources,with a focus on aspects such as network structures and supervision forms,without thoroughly exploring and analyzing the impact of fusion effects on road extraction. This paper proposes an optical image road extraction technique that incorporates fusion strategies.Firstly,an “encoder-decoder” network is adopted as the fundamental structure,and targeted improvements and designs are made based on factors such as the categories and quantities of input data,providing a training and testing framework for subsequent experimental verification. Secondly,to favor the injection of spatial and spectral information,four representative image fusion methods are selected. These methods are utilized to fuse panchromatic and multispectral images,providing the necessary technical support.Finally,the experimental section employs two publicly available datasets to demonstrate the effectiveness of the fusion strategy in improving quantitative evaluation metrics for road extraction. Furthermore,the fusion strategy demonstrates a positive and facilitating effect on the extraction of road areas within challenging regions.

Key words: image fusion, panchromatic image, multispectral image, road extraction, convolutional neural network