The work presented introduces a deep-learning based technique to spectrally unmix data containing more than one endmember. It uses a novel loss function with a soft-attenuation mechanism leading the neural network to focus on visual features of the input spectra. A Deep Neural Network was developed to detect Ammonium Nitrate in Visible to Near Infrared (VNIR) and Short Wave Infrared (SWIR) co-aligned aerial hyperspectral imagery. We compare the target detection accuracy of our method, against a well-known classical method referred to as the Adaptive Cosine Estimation (ACE). We show that our DNN based method outperforms ACE by two-orders of magnitude.