An Interpretable Galaxy Morphology Classification Approach Using Modified SqueezeNet and Local Interpretable Model-agnostic Explanation

Abstract The recent surge in computer vision and deep learning has attracted significant attention within the galaxy morphology community. Various models have been implemented for galaxy morphology prediction with near-perfect accuracy for certain classes. However, many studies treat deep learning models as black-box entities, lacking interpretability of their predictions. To address these limitations while ensuring good performance, we introduced an Improved SqueezeNet (I-SqueezeNet) by incorporating unique residual connections to improve the prediction performance, and we utilize Local Interpretable Model-Agnostic Explanations (LIME) to understand the interpretability. We evaluated the simplified SqueezeNet and I-SqueezeNet, with both channel and vertical concatenation, and compared their performances with those of some exiting methods such as Dieleman’s CNN, VGG13, DenseNet121, ResNet50, ResNext50, ResNext101, DSCNN and customized CNN in classifying galaxy objects using a dataset from the publicly available Galaxy Zoo Data Challenge Project. Our experiments showed that I-SqueezeNet with vertical concatenation achieved the highest average accuracy of 94.08% compared to other methods. Beyond achieving high accuracy, the application of LIME for model interpretation sheds light on the machine learning features and reasoning processes behind the model’s predictions. This information provides valuable insight into the galaxy morphology decision-making process, paving the way for further functional enhancements.