According to EU regulations, countries must monitor noise levels affecting urban populations. Strategic urban scale noise maps are commonly estimated based on physical noise propagation models. Although accurate, this solution is computationally intensive, so the estimation of hourly or even daily noise maps is often not feasible. An alternative approach is to generate surrogate, machine learning-based models for noise map prediction. This solution is less accurate but computationally cheaper. This work aims to analyze the machine learning regression methods and their efficiency and limitations in the task of noise level prediction. The methods were tested on real noise level maps generated for the eastern part of Wrocław, known as Wielka Wyspa. Our results indicate that machine learning-based surrogate noise maps realize the mean accuracy below 2 dB, which seems acceptable in practical applications of urban scale noise maps.