The advent of low-cost acoustic sensor networks in cities raises new interesting approaches for improving the monitoring of the acoustic quality of cities. Many innovative approaches are developed to improve knowledge on sound environments: sound environment recognition, sound source detection, etc. In order to improve the road traffic noise mapping, the use of a specific version of the Non-negative Matrix Factorization (NMF), named thresholded initialized NMF, as a source separation method to estimate the sound level of road traffic from measurements, has proved to be a successful approach. This paper proposes to further detail the functioning of the thresholded initialized NMF on a corpus composed of urban sound scenes mixing traffic and specific interfering components with calibrated sound levels in order to better understand its behavior according to the different sources encountered. The study reveals the different performances of this approach depending on the noise levels of the interfering sources and their proximity to the urban traffic spectrum.