Lightweight and Interpretable Mineral Classification: A Comparative Study of MobileNet Architectures with Grad-CAM and EigenCAM

Reliable and interpretable deep learning models are essential for automated real-world mineral identification, particularly in field or embedded settings where computational resources are strictly limited. This paper presents a comprehensive comparative evaluation of three lightweight convolutional neural network architectures—MobileNetV2, MobileNetV3-Small, and MobileNetV3-Large—for classifying 5,482 images across seven mineral species (biotite, bornite, chrysocolla, malachite, muscovite, pyrite, and quartz) from the Minet v2 dataset. The models are evaluated under a rigorous stratified 3-fold cross-validation scheme with complete per-fold weight reinitialization to prevent data leakage and provide unbiased generalization statistics. To assess model trustworthiness beyond raw predictive accuracy, we systematically compare two explainability techniques, Grad-CAM (gradient-based localization) and EigenCAM (principal component analysis on feature maps), both qualitatively across correctly and incorrectly classified specimens and quantitatively using the Average Drop and Increase in Confidence fidelity metrics. Experimental results yield classification accuracies ranging from 89.3% to 92.0%, with MobileNetV2 achieving the highest overall accuracy (92.01%) and lowest inter-fold variance (0.67%), while MobileNetV3-Large obtains the highest macro F1-score (0.894). Quantitative evaluations consistently demonstrate that Grad-CAM produces more faithful saliency maps than EigenCAM across all evaluated backbones. Crucially, explanation fidelity declines alongside classification accuracy as model size decreases, revealing a notable compactness-explainability trade-off. Furthermore, muscovite and bornite remain the most challenging classes across all architectures, pointing to a data imbalance limitation rather than a structural flaw. These findings offer practical guidance for selecting compact, interpretable models for edge deployment in geological exploration.