This report describes our contribution to the 2016 IEEE AASP DCASE challenge for the acoustic scene classification task. We propose a feature learning approach following the idea of decomposing time-frequency representations with nonnegative matrix factoriza-tion. We aim at learning a common dictionary representing the data and use projections on this dictionary as features for classification. Our system is based on a novel supervised extension of nonnegative matrix factorization. In the approach we propose, the dictionary and the classifier are optimized jointly in order to find a suited representation to minimize the classification cost. The proposed method significantly outperforms the baseline and provides improved results compared to unsupervised nonnegative matrix factorization.