Few-shot crop disease recognition using sequence- weighted ensemble model-agnostic meta-learning.

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Crop diseases threaten global food security, necessitating accurate and timely detection.
• Deep learning for crop disease recognition requires large datasets, which are often unavailable in practice.
• Few-shot learning addresses the challenge of training models with limited data.

Purpose of the Study:

• To introduce a novel few-shot learning approach, SWE-MAML, for training crop disease recognition models with minimal sample sizes.
• To integrate ensemble learning with Model-Agnostic Meta-Learning (MAML) for enhanced few-shot learning performance.

Main Methods:

• Developed the Sequence-Weighted Ensemble Model-Agnostic Meta-Learning (SWE-MAML) framework.
• Employed meta-learning to sequentially train base learners and combined their predictions via weighted summation.
• Integrated ensemble learning within the MAML framework to train multiple classifiers.

Main Results:

• SWE-MAML achieved competitive performance against state-of-the-art algorithms on the PlantVillage dataset.
• SWE-MAML improved accuracy by 3.75%-8.59% compared to the original MAML.
• Optimal performance was observed with 5-7 base learners, and pre-training on more classes improved recognition of unseen classes.
• Achieved 75.71% accuracy on a real-world potato disease recognition task with limited data.

Conclusions:

• SWE-MAML is a highly effective solution for few-shot crop disease recognition, especially in data-scarce agricultural settings.
• The integration of ensemble and meta-learning offers high-performance disease recognition with minimal data.
• SWE-MAML presents a promising approach for practical deployment in precision agriculture.