Physics-Informed Deep Learning for OH(3–1) Airglow Spectra Temperature Retrieval

Abstract. The GRIPS-N instrument maintains one of the longest-running records of mesospheric temperatures, derived from OH(3-1) emission spectra in the near infrared. This study presents two deep neural network architectures for automating the classification and temperature prediction of these spectra, trained on 11 years (2014–2024) of data manually classified by an expert. A dual-branch one-dimensional convolutional classifier separates degraded spectra from those suitable for temperature derivation, achieving a precision of 99.2 % on a test set of 35 456 spectra. A physics-informed temperature model, in which a dense encoder extracts three effective peak intensities that are converted to temperature via the Boltzmann equation, achieves a mean absolute error of 4.17 K and a median error of 2.80 K. Both models use approximately 50 000 trainable parameters and train in minutes on a consumer GPU. The work serves as an informative resource and reference for the application of deep learning in signal processing across a range of remote sensing applications in atmospheric science.