Scholarlink

<p>Sugarcane is a significant commercial crop around the world, but its productivity is severely limited by several insects and pests which have overlapping and frequently coexistent signs of damage. The field diagnosis of pests injury is a challenging task because morphological resemblances between damage caused by borers and other pests often result in diagnostic errors. This study designs and tests a reproducible deep learning pipeline to classify pest damage on sugarcane automatically in images based on convolutional neural networks (CNNs) and ensemble fusion techniques. Both field and controlled condition data were combined to create a consolidated dataset of 4155 expert-annotated images, subdivided into 11 classes (ten different types of pest damage and a control with healthy images). Images were trained and tested on seven pre-trained CNN backbones, AlexNet, ResNet-50, Inception-V3, MobileNet-V2, EfficientNet-B0, DarkNet-53, and YOLO11n-cls, and compared over a fivefold protocol on a Windows-optimized implementation of PyTorch. Of the single networks, EfficientNet-B0 and YOLO11n-cls showed the highest validation accuracies (= 96%), and ResNet-50 scored 97.12 on the last split. Ensemble fusion also displayed robustness: the attention-weighted ensemble reached 96.39% validation accuracy (macro-F1 = 96.18%, ROC–AUC ≈ 0.9985), and the hierarchical super-ensemble had 96.15% accuracy with stable validation dynamics. The models were also exported to TorchScript and ONNX formats, which have been proved to be deployable to real field or mobile applications. Findings showed that a multi-architecture fusion significantly augments generalization of complex, overlapping categories of pests visually. The suggested framework is a scalable and field-capable basis of the implementation of AI-based pest diagnostics into the sugarcane crop management systems.</p>

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

A Reproducible Deep Learning Framework for Field-Deployable Classification of Sugarcane Pest Damage Using Ensemble CNNs

  • 期刊论文

  • 发表日期：

    2026年1月13日

查看全文
  • Sugam Singh,
  • Chandramani Raj,
  • Rajesh U. Modi,
  • Shweta Singh,
  • Arun Baitha,
  • K. Srinivas,
  • V. A. Blessy,
  • S. S. Hasan,
  • R. Viswanathan

摘要

Sugarcane is a significant commercial crop around the world, but its productivity is severely limited by several insects and pests which have overlapping and frequently coexistent signs of damage. The field diagnosis of pests injury is a challenging task because morphological resemblances between damage caused by borers and other pests often result in diagnostic errors. This study designs and tests a reproducible deep learning pipeline to classify pest damage on sugarcane automatically in images based on convolutional neural networks (CNNs) and ensemble fusion techniques. Both field and controlled condition data were combined to create a consolidated dataset of 4155 expert-annotated images, subdivided into 11 classes (ten different types of pest damage and a control with healthy images). Images were trained and tested on seven pre-trained CNN backbones, AlexNet, ResNet-50, Inception-V3, MobileNet-V2, EfficientNet-B0, DarkNet-53, and YOLO11n-cls, and compared over a fivefold protocol on a Windows-optimized implementation of PyTorch. Of the single networks, EfficientNet-B0 and YOLO11n-cls showed the highest validation accuracies (= 96%), and ResNet-50 scored 97.12 on the last split. Ensemble fusion also displayed robustness: the attention-weighted ensemble reached 96.39% validation accuracy (macro-F1 = 96.18%, ROC–AUC ≈ 0.9985), and the hierarchical super-ensemble had 96.15% accuracy with stable validation dynamics. The models were also exported to TorchScript and ONNX formats, which have been proved to be deployable to real field or mobile applications. Findings showed that a multi-architecture fusion significantly augments generalization of complex, overlapping categories of pests visually. The suggested framework is a scalable and field-capable basis of the implementation of AI-based pest diagnostics into the sugarcane crop management systems.