This study aims to enhance the classification of astronomical objects by concentrating on the ResNet50 architecture, which is a deep neural network reorganized for its residual learning abilities. A specialized dataset of twelve space-related categories was collected manually and organized to reflect various image conditions rather than depending on large-scale scientific achievements. We have used various data augmentation methods like rotation, flipping, zooming, and shifting to improve the model's learning and decrease the risk of overfitting. The primary objective was to evaluate the performance of a meticulously fine-tuned ResNet50 model trained on high-variance, visually identical, astronomical or space-based images. Grad-CAM was used to improve interpretability by showing which part of the image influenced the model's classification and generalization. This work helps deepen understanding of how domain-specific tasks, like space photography, where visual patterns often overlap and clarity is limited, can be better supported by standard deep learning models.