基于证据深度学习的小麦病虫害识别模型构建OA

[Objective]This study aims to develop a credible wheat disease and pest identification model to provide technical support for intelligent pest and disease management.[Method]In the inte-lligent recognition of wheat diseases and pests,EfficientNet is selected as an effective feature extrac-tor.A Dirichlet-based framework is utilized to derive uncertainty scores corresponding to model predic-tions,thereby extending the conventional single-class output of EfficientNet to a two-dimensional output comprising both classification results and uncertainty scores.This approach enhances the model's ability to express the trustworthiness of its predictions.Furthermore,the Jaccard index is employed to optimize the consistency between uncertainty score partitioning and classification correct-ness,enabling adaptive uncertainty threshold selection and further improving model credibility and recognition accuracy.[Result]Compared with the traditional EfficientNet,the introduction of the uncertainty mechanism increased the model's precision rate,recall rate,and F1 score by 5.31%,4.42%,and 4.81%,respectively;adding the threshold filtering strategy further improved them by 6.21%,4.33%,and 4.71%;the proposed method achieved performance improvements by 11.84%,8.94%,and 9.74%in the three indicators,respectively.The average uncertainty score for correctly predicted samples is 0.284,which is notably lower than 0.421 for incorrectly predicted ones.With an uncertainty threshold of 0.445 optimized via the Jaccard index,9.51%of high-uncertainty predictions are filtered out,significantly improving the model's credibility in complex environments.[Conclusion]The incorporated uncertainty mechanism allows the model to detect potential misclassifications in com-plex environments and trigger human intervention,thereby significantly enhancing the credibility of decision-making in agricultural applications.