Title: Lessons Learned from the Two Largest Galaxy Morphological Classification Catalogues built by Convolutional Neural Networks

Abstract: We compare the two largest galaxy morphology catalogues, which separate early and late type galaxies at intermediate redshift. The two catalogues were built by applying supervised deep learning (convolutional neural networks, CNNs) to the Dark Energy Survey data down to a magnitude limit of $\sim$21 mag. The methodologies used for the construction of the catalogues include differences such as the cutout sizes, the labels used for training, and the input to the CNN - monochromatic images versus $gri$-band normalized images. In addition, one catalogue is trained using bright galaxies observed with DES ($i<18$), while the other is trained with bright galaxies ($r<17.5$) and `emulated' galaxies up to $r$-band magnitude $22.5$. Despite the different approaches, the agreement between the two catalogues is excellent up to $i<19$, demonstrating that CNN predictions are reliable for samples at least one magnitude fainter than the training sample limit. It also shows that morphological classifications based on monochromatic images are comparable to those based on $gri$-band images, at least in the bright regime. At fainter magnitudes, $i>19$, the overall agreement is good ($\sim$95\%), but is mostly driven by the large spiral fraction in the two catalogues. In contrast, the agreement within the elliptical population is not as good, especially at faint magnitudes. By studying the mismatched cases we are able to identify lenticular galaxies (at least up to $i<19$), which are difficult to distinguish using standard classification approaches. The synergy of both catalogues provides an unique opportunity to select a population of unusual galaxies.