A benchmark analysis of saliency-based explainable deep learning methods for the morphological classification of radio galaxies

TL;DR

This paper evaluates and compares 10 saliency-based explainability methods for deep learning models in radio galaxy classification, emphasizing interpretability and identifying the most meaningful attribution techniques.

Contribution

It introduces a framework for assessing explainability methods in astronomy, highlighting the effectiveness of Score-CAM, Grad-CAM, and Grad-CAM++ in producing astrophysically relevant attributions.

Findings

Score-CAM, Grad-CAM, and Grad-CAM++ produce meaningful attribution maps

Other methods often highlight irrelevant or noisy regions

The framework helps evaluate interpretability in astronomical deep learning models

Abstract

This work proposes a saliency-based attribution framework to evaluate and compare 10 state-of-the-art explainability methods for deep learning models in astronomy, focusing on the classification of radio galaxy images. While previous work has primarily emphasized classification accuracy, we prioritize model interpretability. Qualitative assessments reveal that Score-CAM, Grad-CAM, and Grad-CAM++ consistently produce meaningful attribution maps, highlighting the brightest regions of FRI and FRII galaxies in alignment with known astrophysical features. In contrast, other methods often emphasize irrelevant or noisy areas, reducing their effectiveness.