Scholarlink

Coconut Rhinoceros Beetle (CRB) infestation represents a major agricultural challenge, particularly in tropical regions where coconut plantations are vital to economies. Traditional detection methods rely on manual inspections, which are labor-intensive and prone to delays, leading to widespread crop damage. This study builds on meta-learning paradigms to address these limitations, adapting Relation Networks, a technique that learns similarity metrics between support and query samples, for efficient classification of infested coconut tree crowns. CRB causes distinctive V-shaped notches on palm fronds, but visual similarities between healthy and infested samples complicate detection. Exploratory analysis, including RGB value comparisons and Shannon entropy, showed no significant differences via paired t-tests, showing the need for sophisticated feature extraction. Dataset imbalance (40% healthy vs. 60% infested) and variability from drone-captured images further exacerbate challenges in achieving high accuracy with conventional classifiers. Images are preprocessed using YOLO bounding boxes to isolate crown regions. We evaluate lightweight CNNs (ResNet18, ShuffleNet, SqueezeNet) and propose NotchNet, an attention-enhanced model with CBAM for focused feature extraction, outputting 256-dimensional embeddings. These embeddings feed into Relation Networks, which compute relational scores between support prototypes and queries via a learned relation module. Trained on the Roboflow dataset (2,579 train, 249 validation, 133 test images) with augmentations like flipping and brightness adjustments, the framework uses episodic sampling (C classes, K supports, Q queries) and optimizes with Adam and cosine annealing. NotchNet with Relation Networks achieved 83.0% accuracy, 0.85 precision, 0.84 recall, and 0.84 F1-score in few-shot tasks, outperforming the baseline Prototypical Networks (52.9% accuracy). When paired with traditional classifiers, NotchNet + Decision Trees reached 65.4% accuracy, while SqueezeNet + SVC recorded 71.0% accuracy. The proposed pipeline also demonstrated the lowest inference time of 0.0005 s per image, which shows its suitability for real-time edge deployment.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Few-Shot Classification of Coconut Rhinoceros Beetle Infestation in Drone Images Using Relation Networks

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • G. Anusuya Devi,
  • Ashwinth Janarthanan

摘要

Coconut Rhinoceros Beetle (CRB) infestation represents a major agricultural challenge, particularly in tropical regions where coconut plantations are vital to economies. Traditional detection methods rely on manual inspections, which are labor-intensive and prone to delays, leading to widespread crop damage. This study builds on meta-learning paradigms to address these limitations, adapting Relation Networks, a technique that learns similarity metrics between support and query samples, for efficient classification of infested coconut tree crowns. CRB causes distinctive V-shaped notches on palm fronds, but visual similarities between healthy and infested samples complicate detection. Exploratory analysis, including RGB value comparisons and Shannon entropy, showed no significant differences via paired t-tests, showing the need for sophisticated feature extraction. Dataset imbalance (40% healthy vs. 60% infested) and variability from drone-captured images further exacerbate challenges in achieving high accuracy with conventional classifiers. Images are preprocessed using YOLO bounding boxes to isolate crown regions. We evaluate lightweight CNNs (ResNet18, ShuffleNet, SqueezeNet) and propose NotchNet, an attention-enhanced model with CBAM for focused feature extraction, outputting 256-dimensional embeddings. These embeddings feed into Relation Networks, which compute relational scores between support prototypes and queries via a learned relation module. Trained on the Roboflow dataset (2,579 train, 249 validation, 133 test images) with augmentations like flipping and brightness adjustments, the framework uses episodic sampling (C classes, K supports, Q queries) and optimizes with Adam and cosine annealing. NotchNet with Relation Networks achieved 83.0% accuracy, 0.85 precision, 0.84 recall, and 0.84 F1-score in few-shot tasks, outperforming the baseline Prototypical Networks (52.9% accuracy). When paired with traditional classifiers, NotchNet + Decision Trees reached 65.4% accuracy, while SqueezeNet + SVC recorded 71.0% accuracy. The proposed pipeline also demonstrated the lowest inference time of 0.0005 s per image, which shows its suitability for real-time edge deployment.