Addressing representation instability, strategy rigidity, and class imbalance in active learning for agricultural pest and disease recognition(解决农业病虫害识别主动学习中表征不稳定、策略僵化和类别不平衡问题)

Publication - Article

Authors

2. Corresponding Author

Abstract

Agricultural pests and diseases cause substantial crop losses worldwide, yet deploying deep learning in smart agriculture is hindered by the high cost of expert annotation. Generative AI (GAI) techniques such as GANs can augment training data, but generated samples still require expert verification and their quality depends on real labeled data, leaving the annotation bottleneck unresolved. Active learning offers a complementary solution by selecting the most informative samples for labeling, yet existing methods suffer from three limitations: representation instability due to static prototype computation on noisy field images, strategy rigidity from fixed selection ratios across training stages, and neglect of class imbalance during sample selection. We propose a three-level active learning framework that introduces targeted innovations at each level of the learning pipeline. Dynamic Prototype Memory maintains stable class representations through momentum-based exponential moving average updating, providing provably lower variance than static prototype recomputation. Context-Aware Sample Selection employs a neural contextual bandit to dynamically adjust the branch combination ratio at each round, learning how to combine complementary selection strategies rather than which individual samples to select. Class-Balanced Scoring incorporates inverse-frequency weighting to prioritize underrepresented classes directly at the selection stage, filling a gap left unaddressed by existing active learning methods. Experiments on IP06 (insect pest) and Natural CW-10 (weed) datasets demonstrate that the framework achieves 83.36% and 85.07% accuracy at 90% data availability, outperforming the strongest competitor by 4.12 percentage points on IP06 and 2.68 percentage points on Natural CW-10 at 20% data availability, with the largest gains concentrated in low-data regimes. Integration with GAN-augmented data yields further additive improvements, with combined accuracy reaching 84.91% on IP06 and 86.42% on Natural CW-10, confirming the complementary role of active learning within GAI-empowered smart agriculture pipelines. These results demonstrate that principled annotation selection can substantially reduce labeling costs while maintaining competitive recognition performance, offering a practical and scalable solution for real-world smart agriculture deployment where expert annotation resources are severely constrained. Ablation studies verify that the three innovations are mutually complementary, and sensitivity analysis shows robustness to hyperparameter variation.

Acknowledgements

Declaration of competing interest The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. Acknowledgments This work was supported by Start-up Research Fund of Tianjin University of Technology (No. 01002405).

Publication references - 23

Sorted by: Date

2025, Computers and Electronics in Agriculture - Article

2025, Computers and Electronics in Agriculture - Article

2025, IEEE Transactions on Neural Networks and Learning Systems - Article

2024, Applied Soft Computing - Article

2024, Crop Protection - Article