The modelling of auditory scenes is a challenging task in Computational Auditory Scene Analysis. A method based on sparse Non-negative Matrix Factorization that can be used with no prior knowledge of the audio content to establish the similarity between scenes is proposed. The method is evaluated on a corpus of soundscapes of train stations issued from a perceptual study and results are compared with the human perception. The proposed method, by being able to focus on salient events within the scene, achieves better performances than a state-of-the-art Bag-of-Frames approach though not reaching the human performances.