DCMF: Deep Counterfactual Metric Framework for limited data plant disease recognition

Deep learning methods have achieved remarkable success in plant disease recognition. However, these methods rely on large-scale labeled datasets for training to ensure the reliability of empirical risk minimization. In the real world, obtaining such extensive disease data remains challenging. With limited data, traditional correlation-based learning frameworks could establish spurious correlations between disease data and disease classes, which severely harms their generalization ability. We address this issue from a causal perspective by proposing the Deep Counterfactual Metric Framework (DCMF). Specifically, DCMF employs a Counterfactual Reasoning Module (CRM) to construct a counterfactual world where each disease image contains only healthy features, enabling estimation of the direct effect of healthy regions on disease recognition. By subtracting this direct effect from the total effect on classes, we effectively eliminate spurious correlations, allowing the model to learn robust disease-specific features for reliable generalization in limited data scenarios. Extensive experiments on PlantVillage and PlantLeaves datasets under 5-shot and 10-shot settings demonstrate that DCMF achieves an average performance improvement of 7.2% over the best baseline methods. These improvements validate the effectiveness of DCMF in limited data plant disease recognition. • To the best of our knowledge, we are the first to construct a structural causal model in limited data scenarios based on causal inference for plant disease recognition. • Based on a constructed counterfactual world where disease images contain only healthy features, we propose a Deep Counterfactual Metric Framework (DCMF) that eliminates the spurious correlations between disease samples and disease classes via the proposed Counterfactual Reasoning Module (CRM). • We conduct extensive experiments on two datasets, demonstrating significant improvements over baseline methods. Additionally, we integrate the CRM into four baseline methods and achieve better results, which illustrates that the CRM is effective.