Given the prominent local government debt issues in China, effectively evaluating infrastructure benefits and preventing over investment has become critically important. We propose a scalable method that leverages high-resolution satellite images and nighttime light data to address the challenges of data scarcity and lag in rapidly urbanizing emerging towns. Specifically, we improve the DeepGlobe Challenge first-place road extraction model with a non-local block and a patch refinement module, achieving a 10% improvement in intersection-over-union (IoU) over the baseline. Using road density data for 544 Chinese counties from 2015 to 2021, we find that a 1% increase in road density corresponds to a 0.28% increase in nighttime light intensity, with the strongest effects observed in counties with moderate population, service provision, or fiscal capacity. These results offer important insights for smart city planning and infrastructure investment allocation.