Object detection methods based on deep learning have significantly reduced time-consuming tasks. Semantic segmentation has shown remarkable progress in the study of rocks, especially when applied to petrographic thin sections. Despite the development of various models for specific applications in this field with promising results, their widespread adoption remains limited. This hesitation is largely due to a lack of user confidence stemming from the absence of explainability in the outcomes provided by these models. This study explores the explainability of the state-of-the-art YOLOv11 model in detecting andalusite, biotite, and grains with oolitic textures. We trained three models using plane-polarized-light thin-section microphotographs of the selected targets. Subsequently, we applied color and singular value perturbations to the annotated images using color masks and analyzed the model's inference through connected region heatmaps. Our findings suggest that the trained models prioritize low-frequency attributes like shape, predominant colors, and contrast over the studied targets' characteristic tones. These insights contribute to the practical application of deep learning for detecting and segmenting grains and minerals in thin sections.