An Evaluation of XAI Methods for Object Detection in Satellite Images using YOLOv5

Özenç U., Ertürk A.

Fifth International Congress of Applied Statistics (UYIK-2024), İstanbul, Türkiye, 21 - 23 Mayıs 2024, ss.1-9

  • Yayın Türü: Bildiri / Tam Metin Bildiri
  • Basıldığı Şehir: İstanbul
  • Basıldığı Ülke: Türkiye
  • Sayfa Sayıları: ss.1-9
  • Kocaeli Üniversitesi Adresli: Evet

Özet

In recent years, deep learning based approaches have gained widespread adoption in Earth observation and remote sensing, mirroring their success in numerous other domains. However, unlike approaches based on physical models, deep learning methods operate as black boxes, concealing internal processes influencing final decisions. This lack of transparency poses a challenge, particularly in applications where interpretability is paramount, as outputs generated by these approaches cannot be fully trusted or verified. Explainable Artificial Intelligence (XAI) aims to make the deep learning processes and their outputs more interpretable for researchers and end users. The purpose of this study is to investigate and evaluate the performance of various XAI methodologies for post-hoc explainability of object detection in satellite images using deep learning. Class-activation mapping (CAM) based XAI methods, namely GradCAM, GradCAM++, EigenCAM, ScoreCAM, and LayerCAM, are used for post-hoc explainability, following target detection by You Look Only Once (YOLO) algorithm. Experimental results show that each method provides considerably different saliency maps, which may be used for qualitative performance analysis of the interpretability provided by these methods. However, in a large dataset, a qualitative analysis by itself may be subjective and misleading. As such, an evaluation framework tailored for remote sensing applications is adopted to evaluate the interpretability performances of these XAI methods quantitatively. The findings provide an important step towards understanding the role and effectiveness of these XAI methods for interpretability of object detection for remote sensing.