Explainable artificial intelligence for sedimentary facies segmentation.

Area of Science:

• Geosciences
• Artificial Intelligence
• Sedimentology

Background:

• Artificial intelligence (AI) is increasingly transforming geosciences.
• Explainable AI (XAI) models are crucial for scientific research and practical adoption.
• Accurate sedimentary facies identification is vital for geological interpretation.

Purpose of the Study:

• To develop and evaluate a general yet tailored pipeline for semantic segmentation of sedimentary facies from Holocene core images.
• To enhance model interpretability using explainable AI techniques.
• To demonstrate the value of XAI in geosciences for transparent decision-making.

Main Methods:

• Evaluation of convolutional neural networks (CNNs) and Transformer-based backbones for optimal architecture.
• Systematic patch-size ablation to align model reasoning with expert sedimentological knowledge.
• Application of saliency mapping (grad-CAM) and predictive entropy for enhanced interpretability and uncertainty quantification.

Main Results:

• The developed explainable AI framework significantly advances automatic facies prediction.
• The pipeline provides actionable insights into model behavior and geological relevance.
• Saliency mapping and predictive entropy effectively highlight meaningful geological regions and quantify uncertainty.

Conclusions:

• Explainable AI pipelines offer substantial added value in geosciences.
• Transparent decision-making in AI is essential for bridging research and real-world applications in geology.
• This approach facilitates the integration of AI in critical geoscience domains.