<p>Supervised machine learning and deep learning algorithms have been commonly applied in identifying geochemical anomalies related to mineralization. However, the performance of supervised models is often hindered by two critical challenges of training deposit-type locations: (i) the scarcity of known mineral deposits available for training and (ii) the heterogeneity within multivariate spatial data signatures of deposits. While previous studies have attempted to address these issues through various data augmentation techniques and sample refinement strategies, effectively learning from limited and heterogeneous samples remains a significant hurdle. To address these limitations, this paper proposes a novel method that synergistically integrates unsupervised clustering algorithms, namely K-means and hierarchical density-based spatial clustering of applications with noise (HDBSCAN), with model-agnostic meta-learning (MAML). The proposed framework employs unsupervised clustering to partition the heterogeneous deposit-type locations samples into distinct, geochemically consistent subgroups, thereby addressing the issue of sample variability. Subsequently, a cluster-stratified sampling strategy is applied to construct representative few-shot tasks from these subgroups, ensuring that the full spectrum of sample characteristics is captured while preventing model bias. Experimental results demonstrated that this approach significantly outperforms conventional convolutional neural network models. Specifically, the non-augmented HDBSCAN-stratified MAML (HS-MAML-NA) model proved most effective, achieving the highest predictive accuracy with an AUC of 0.868 in predicting 21 deposits. Consequently, this integrated approach enhances the robustness and predictive accuracy in identifying geochemical anomalies related to mineralization, providing a more effective tool for mineral exploration, especially in geologically complex regions with sparse mineral deposits.</p>