Spectral feature extraction is a crucial procedure in automated spectral analysis. In this paper, we present a new automated feature extraction method for astronomical spectra, with application in LAMOST spectral classification and defective spectra recovery. The basic idea of our approach is to train a deep neural network to extract features of spectra with pseudo-inverse learning algorithm in an analytical way without any iterative optimization procedure and any control parameters tuning. The proposed method can be regarded as a new valid alternative general-purpose feature extraction method for various tasks in spectral data analysis.

Defective spectra with their corresponding recoveries. The spectra marked in red are the synthetic defective ones. The black spectra are the ones repaired with our method.

Visualization of the importance of input patterns for the predicted results in three randomly selected spectra with different types. We systematically cover up different portions of the input spectra with a sliding window to see how the final classifier output changes. Rows 1 to 3 show the three sampled spectra separately. The size of the sliding window is specified as 1, 64, 128 and 256, respectively, in columns 1 to 4