In this study 300 skin lesion (including 32 skin melanomas) multispectral data cubes were analyzed. The multi-step and single step machine learning approaches were analyzed to find the wavebands that provide the most information that helps discriminate skin melanoma from other benign pigmented lesions. The multi-step machine learning approach assumed training several models but proved itself to be ineffective. The reason for that is a necessity to train a segmentation model on a very small dataset and utilization of standard machine learning classifier which have shown poor classification performance. The single-step approach is based on a deep learning neural network. We have conducted 2600 experiments on two neural network architectures: Popular pre-trained image analysis "InceptionV3" and simple custom convolutional neural network (ConvNet) classifiers. Observing performance metrics of these two deep-learning (DL) based architectures allowed to determine combinations of three spectral wavebands allowing to train a classifier with the best classification results. It was found that a simple ConvNet classifier allowed us to get better classification results. ConvNet training results have shown that most informative wavebands are 450nm which is the most informative for melanin concentration on the skin surface, 590nm that represents integral information about melanin and hemoglobin distribution from epidermis-dermis layer, and 950 nm that provide information from deeper skin layers. As introduced the convolutional neural network (CNN) model was simple but has not shown great performance. Also, we have to explore alternative CNN architectures. AutoKeras framework was used to find an architecture of the image classifier using the found waveband triplets.