Few-Shot Action Recognition and Explainable Learning for Agricultural Waste Disposal

To address the challenges of limited sample size, data heterogeneity, and lack of interpretability in agricultural waste disposal behavior recognition, this study proposes the MAGTfmr framework. The framework integrates meta-learning, CGAN-based data augmentation, and prototypical networks to construct a few-shot recognition model, while a Transformer is employed to achieve cross-modal mapping between actions and textual descriptions. Finally, SHAP analysis is used to quantify the marginal contributions of multi-source variables to behavioral decisions. Results show that MAG-Tfmr outperforms various baseline models in terms of recognition accuracy, generalization, and semantic interpretability. Moreover, the explainable learning outcomes reveal that key variables-such as action types, tool usage, and environmental awareness-have significant driving effects on behavioral choices.