Conventional multi-modal fusion approaches assume that quality is uniformly distributed across samples, overlooking dynamic inter-modal reliability. Such static strategies struggle to down-weight low-quality modalities when data heterogeneity is high, rendering the fused representation susceptible to noise, missing modalities, and other degradations, thereby diminishing fusion benefits. Under small-sample conditions, these limitations further erode classifier robustness. To enhance reliability and adaptability in small-sample texture recognition, we propose the small-sample multimodal evidence fusion framework for texture classification (SMEF-TC). Built on subjective logic, SMEF-TC leverages a Dirichlet distribution to jointly model class probabilities and epistemic uncertainty, thereby eliminating extra uncertainty-quantification during inference and the high computational overhead of traditional Bayesian methods. Incorporating modality-specific uncertainties, evidence theory fuses multi-modal information, enabling the model to adaptively recalibrate each modality′s contribution and effectively suppress redundant or noisy cues. By simultaneously accounting for information evidence and predictive uncertainty, SMEF-TC retains high recognition accuracy under conditions of noise, modality absence, and imbalanced quality. Experiments on the public LMT-108 and LMT-184 texture datasets yield accuracies of 96.53% and 94.70%, respectively, confirming that SMEF-TC offers superior precision and robustness for small-sample texture classification compared with existing techniques.