Deep Learning for Classification of Low Surface Brightness Galaxies in Dark Energy Survey

Low Surface Brightness Galaxies (LSBGs) are faint, diffuse systems that are difficult to identify and are often confused with imaging artifacts in wide-field surveys. In this work, we apply convolutional neural networks (CNNs) to classify LSBGs in the full six-year Dark Energy Survey (DES Y6) dataset, extending earlier analyses on the Year 3 data. Training on the $\sim$40,000 labeled Y6 objects used in prior work, our CNN achieves 90.7\% accuracy, surpassing traditional feature-based methods. Applied to the full Y6 sample of $\sim$70,000 labeled-objects, the network reliably identifies artifacts, but struggles with ambiguous human-labeled LSBG cases. Targeted augmentation reduces the fraction of LSBGs classified as artifacts from 27.4\% to 18.5\%, bringing CNN predictions into closer alignment with human labels. We also compare two classifiers that highlight a trade-off: one favors completeness by retaining more LSBG candidates (at the risk of inaccuracy), while the other favors purity by excluding ambiguous cases. Overall, CNNs classify LSBGs with high efficiency and accuracy while also uncovering potential human mislabels. With improved training data and stronger architectures, CNN-based approaches will be indispensable for understanding the low-surface-brightness universe in future large-scale surveys.