<p>Recent advances in artificial intelligence are transforming geosciences, with growing attention on explainable models to support both scientific research and practical adoption. Here, we present a general yet tailored pipeline for semantic segmentation of sedimentary facies from Holocene core images. We first identify the optimal model architecture, evaluating both convolutional neural networks and Transformer-based backbones, then perform a systematic patch-size ablation to better align with expert-level sedimentological reasoning. Finally, we enhance interpretability through saliency mapping with grad-CAM and predictive entropy, which highlight geologically meaningful regions and quantify model uncertainty. Our results demonstrate that an explainable AI framework not only takes a significant step forward in automatic facies prediction but also provides actionable insights into model behavior. This approach emphasizes the added value of explainable pipelines in critical domains of geosciences, where transparent decision-making is essential for bridging research and real-world applications.</p>