F1 scores of CA U-Net, SVM, and OT models.

The extraction of roadways from remote sensing imagery constitutes a pivotal task, with far-reaching implications across diverse domains such as urban planning, management of transportation systems, emergency response initiatives, and environmental monitoring endeavors. Satellite images captured during daytime have customarily served as the primary resource for this extraction process. However, the emergence of Nighttime Light (NTL) remote sensing data introduces an innovative dimension to this arena. The exploration of NTL data for road extraction remains in its nascent stage, and this study seeks to bridge this gap. We present a refined U-Net model (CA U-Net) integrated with Cross-Attention Mechanisms, meticulously designed to extract roads from Yangwang-1 NTL images. This model incorporates several enhancements, thereby improving its proficiency in identifying and delineating road networks. Through extensive experimentation conducted in the urban landscape of Wenzhou City, the model delivers highly accurate results, achieving an F1 score of 84.46%. These outcomes significantly surpass the performance benchmarks set by Support Vector Machines (SVM) and the Optimal Threshold (OT) method. This promising development paves the way towards maximizing the utility of NTL data for comprehensive mapping and analysis of road networks. Furthermore, the findings underscore the potential of utilizing Yangwang-1 data as a reliable source for road extraction and reaffirm the viability of deploying deep learning frameworks for road extraction tasks utilizing NTL data.