Grad-CAM-Based Fine-Grained Classification for Celestial Object

This paper proposes a new method of classifying celestial objects using Grad-Class Activation Mapping (Grad-CAM) in conjunction with deep learning models such as VGG16, Google Net, Dense Net, and Efficient Net. The paper looks at the difficulties characterizing the enormous varieties of astronomical objects and the complexity that makes it difficult to classify the objects by conventional methods. From the empirical evaluation, it is clear that the Dense Net model has the ability to predict the right class of celestial objects with the help of Grad-CAM and has the highest accuracy of 99.76% for differentiating between images in the case of nebula. Grad-CAM works well at identifying the features that affect the model’s output. Thus, the integration of Grad-CAM with deep learning helps to increase the classification indicators and offers useful information regarding the morphological and spectral properties of the objects, thus, contributing to the further development of observational astronomy. It outperformed Class Activation Mapping (CAM) in terms of performance.