Integrating SHAP Explainability in Few-Shot Lithology Identification Using Dynamic Semi-Supervised Meta-Learning

Lithology classification is a crucial task in geological exploration, playing a significant role in oil and gas exploration as well as mineral resource development. However, traditional supervised learning methods rely heavily on large amounts of labeled data, and obtaining labeled well-logging data is both costly and time-consuming, significantly limiting their widespread application. To address this issue, this paper proposes a dynamic semi-supervised meta-learning with SHAP (DSSMLS) method, which achieves efficient and accurate lithology classification under limited labeled data conditions. DSSMLS adopts a meta-learning framework to enable rapid generalization using only a small number of labeled samples. It further integrates a semi-supervised learning strategy to leverage unlabeled data and enhance the model’s generalization ability. To mitigate the error accumulation issue commonly associated with traditional pseudo-labeling methods, DSSMLS incorporates a dynamic pseudo-label generation and prototype correction mechanism, which adaptively refines class prototypes to improve the stability of classification decisions. Additionally, the model integrates attention mechanisms, to enhance feature extraction from well-logging data. To improve model interpretability, DSSMLS combines SHAP (SHapley Additive ExPlanations) analysis to quantify the influence of key well-logging parameters on classification decisions. This study conducts experiments using well-logging data from the Tarim Basin oilfield in China. Experimental results demonstrate that DSSMLS significantly outperforms baseline models.