Parameter measurement based on photometric images

The gas-phase metallicity is a crucial parameter for understanding the evolution of galaxies. Considering that the number of multiband galaxy images can typically reach tens of millions, using these images as input data to predict gas-phase metallicity has become a feasible method. However, the accuracy of metallicity estimates from images is relatively limited. To solve this problem, we propose the galaxy parameter measurement residual network (GPM-ResNet), a deep learning method designed to predict gas-phase metallicity from photometric images of DESI. The parameters of photometric images are labeled with gas-phase metallicity values, which were obtained through spectroscopic methods with a high accuracy. These labeled images serve as the training dataset for the GPM-ResNet method. GPM-ResNet mainly consists of two modules: a multi-order feature extractor and a parameter generator, enhancing the ability to effectively extract features related to gas-phase metallicity from photometric images. The σ of Zpred – Ztrue is 0.12 dex, which significantly outperforms the predicted results of the second-order polynomial (σ=0.16 dex) and the third-order polynomial (σ=0.16 dex) fit using the color-metallicity relation on the same dataset. To further emphasize the superiority of GPM-ResNet, we analyzed the predicted results on various network architectures, galaxy sizes, image resolutions, and wavelength bands of images. Moreover, we explored the mass-metallicity relation and recovered the relation successfully by utilizing the predicted values, Zpred. Finally, we applied GPM- ResNet to predict the gas-phase metallicity of spiral (EXP) galaxies observed by DESI, resulting in a comprehensive catalog containing 5 095 815 pieces of data.