Neural-Based Compression for the Spectral Data of the New Vacuum Solar Telescope

Due to the rapid increase in spectral data generation as well as storage and transmission constraints, data compression has become particularly important for the New Vacuum Solar Telescope (NVST) at Yunnan Observatory. In this paper, we present a method for compressing NVST Ca II (8542 Å) spectral data based on a Convolutional Variational Autoencoder (VAE). Our results show that the compression ratios of the VAE-based approach may achieve as high as 107, while keeping the error between the decompressed data and the original data within the inherent error range of the raw data. This is much better than the appropriate compression ratio of 30 that is attained using the current PCA-based approach. Furthermore, the stability of the VAE approach is demonstrated by the almost constant differences between the VAE-compressed data and the raw data when the compression ratio ranges from 8 to 107. We also investigated Doppler velocity images deduced from the VAE-compressed data and found that the error in Doppler velocity is significantly less than 5 km s⁻¹ when the compression ratio does not exceed 107.