Hyperspectral images provide fine details of the observed scene from the exploitation of contiguous spectral bands. However, the high dimensionality of hyperspectral images causes a heavy burden on processing. Therefore, a common practice that has been largely adopted is the selection of bands before processing. Thus, in this work, a new unsupervised approach for band selection based on autoencoders is proposed. During the training phase of the autoencoder, the input data samples have some of their features turned to zero, through a masking noise transform. The subsequent reconstruction error is assigned to the indices with masking noise. The bigger the error, the greater the importance of the masked features. The errors are then summed up during the whole training phase. At the end, the bands corresponding to the biggest indices are selected. A comparison with four other band selection approaches reveals that the proposed method yields better results in some specific cases and similar results in other situations.