The concept of joint communications and sensing (JCAS) enables a wireless network to sense its environment. This means in particular that the network can perceive objects that influence the propagation of transmitted signals, which opens up the possibility to improve the construction of radio environment maps (REMs). REMs are an essential tool for spectrum monitoring which becomes more and more important as the spectrum is a bottleneck in today's wireless networks. The paper proposes a machine learning (ML)-based approach that combines knowledge from a distributed sensor network and knowledge on obstacles to create an REM without requiring knowledge on the transmitter location. The proposed approach is evaluated and compared against two other methods based on simulated data for different sensor grid sizes. In the case of a sparse sensing network, the approach outperforms Kriging as well as an ML-based approach that only uses received power data. An outlook on further research in the described direction is provided at the end.