Recently the 'soft' classification approach has gained in popularity against the discrete (hard) classification of land cover from remotely-sensed imagery. An empirical study is presented to test training procedures with neural networks for soft (mixture) classification. The results show thatland cover mixtures are best recognized following training with two-component mixed pixels, and that linearly re-scaled or binned target vector representations are equally satisfactory. Interestingly, dominant classes within pixels are also better recognized by training with wider varieties of class mixtures.