Evaluating the spatial inclination of disk galaxies with Convolutional Neural Network

The inclination of spiral galaxies plays an important role in their astronomical analyses, such as measuring their distances using the correlation between absolute luminosity and rotation rate. Improving the accuracy of distance measurements reduces the uncertainty of inferred peculiar motions, thereby improving the precision of the large-scale structure map of the local universe generated in programs such as Cosmicflows. The inclinations of a disk galaxy can be roughly derived from the axial ratio of the projected ellipse that defines its boundary. However, this approximation only provides good enough inclination estimates in ~30% of cases for various reasons, such as the existence of prominent bulges that dominate the axial ratio measurements. In this study, we build multiple models by employing the Convolutional Neural Network (CNN) of different architectures to determine the inclination of spiral galaxies from their visible images. We utilize ~20,000 galaxy images, taken from the SDSS image archive, to train and evaluate these models. The inclinations of all of our sample galaxies have been manually determined with the collaboration of citizen scientists. Exploring different CNN structures, we found that models with the convolutional filters of size 3×3 are simple but yet powerful to meet our requirements, considering the limited computational resources. We report that averaging across multiple training scenarios and model architectures improves the overall accuracy of predictions. Our trained networks can measure inclinations with the root-mean-square uncertainty of ~3 degrees, which is comparable to the average human performance, i.e. ~2.6 degrees. All of our studied models exhibit better accuracy than the ellipticity-based formalism.