Spectral image classification with deep learning

Abstract

We present the Spectral Image Typer (SPIT), a convolutional neural network (CNN) built to classify spectral images. In contrast to traditional, rules-based algorithms that rely on metadata provided with the image (e.g., header cards), SPIT is trained solely on the image data. We have trained SPITon 2004 human-classified images taken with the Kast spectrometer at Lick Observatory with types of Bias, Arc, Flat, Science, and Standard. We include several preprocessing steps (scaling, trimming) motivated by human practice and also expanded the training set to balance between image type and increase diversity. The algorithm achieved an accuracy of 98.7% on the held-out validation set and an accuracy of 98.7% on the test set of images. We then adopt a slightly modified classification scheme to improve robustness at a modestly reduced cost in accuracy (98.2%). The majority of mis-classifications are Science frames with very faint sources confused with Arc images (e.g., faint emission line galaxies) or Science frames with very bright sources confused with Standard stars. These are errors that even a well-trained human is prone to make. Future work will increase the training set from Kast, will include additional optical and near-IR instruments, and may expand the CNN architecture complexity. We are now incorporating SPIT in the PYPIT data reduction pipeline (DRP) and are willing to facilitate its inclusion in other DRPs.