CropCLR-Wheat: A Label-Efficient Contrastive Learning Architecture for Lightweight Wheat Pest Detection

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Field-based wheat pest recognition faces challenges like viewpoint variations, limited data, and diverse data distributions.
• Accurate and robust pest identification is crucial for effective crop management and yield optimization.

Purpose of the Study:

• To develop an advanced pest identification framework, CropCLR-Wheat, addressing current limitations in wheat pest recognition.
• To improve the accuracy, robustness, and efficiency of automated pest detection systems in agriculture.

Main Methods:

• Integration of self-supervised contrastive learning with an attention-enhanced mechanism.
• Implementation of a viewpoint-invariant feature encoder and a diffusion-based feature filtering module.
• Evaluation using 5-shot and 10-shot learning, semantic segmentation, and robustness tests under viewpoint disturbances.

Main Results:

• Achieved high precision (89.4%-92.3%) and recall (87.1%-90.5%) in limited-sample classification tasks.
• Demonstrated superior performance in semantic segmentation (mIoU 82.7%) compared to SegFormer and Mask R-CNN.
• Exhibited strong robustness with a prediction consistency rate of 88.7% and efficient deployment on edge devices (84 ms latency, 11.9 FPS).

Conclusions:

• CropCLR-Wheat effectively overcomes challenges in field-based wheat pest recognition, offering high accuracy and robustness.
• The framework's balance of generalization performance and deployability supports intelligent agricultural systems.
• The proposed method shows significant potential for widespread application in precision agriculture and edge computing environments.