Towards Explainable Discriminative Models for EO Data Analysis

Explainability in deep learning is a crucial aspect of ensuring transparency, trust, and interpretability in model predictions, particularly in complex applications like satellite imagery analysis. As deep learning models become more intricate, their decision-making processes become increasingly opaque, making it difficult for researchers and practitioners to assess how and why certain predictions are made. Post-hoc explainability techniques aim to provide insights into these black-box models by analyzing their decisions after training. This chapter explores post-hoc methods applied to deep learning models, including segmentation networks, graph-based spatial modeling using Graph Convolutional Networks, and temporal analysis with Recurrent Neural Networks. We also discuss how these techniques can help identify biases, improve model reliability, and enhance decision-making in satellite image analysis. Additionally, we highlight existing challenges in post-hoc explainability and explore potential directions for future research to develop more transparent and robust AI systems. Understanding why a segmentation model classifies certain pixels as roads or buildings, or how a graph-based model determines spatial relationships between objects, is critical for validating results and making informed decisions. Post-hoc explainability not only improves model trustworthiness but also assists in optimizing models by revealing weaknesses in predictions, guiding refinements in training strategies, and enabling better human-AI collaboration. This chapter provides a structured approach to applying explainability techniques to spatial and temporal modeling, offering a comprehensive understanding of their implications for real-world applications.