Scholarlink

<p>Whiteflies are one of the most common and harmful pests in agriculture, capable of rapid spreading and severe economic losses. In automated whitefly detection, limited annotated data, background variances, and shifting lighting conditions hinder model generalization. The key contribution of this study lies in the implementation of an intermediate agricultural domain adaptation stage prior to final fine-tuning on the target dataset. The proposed sequential transfer learning approach commences with a deep learning You Only Look Once (YOLOv8) model pretrained on the MSCOCO dataset. Next, the model is exposed to the publicly available Cassava dataset (3,000 whitefly images). This enables intermediate-domain adaptation and hyperparameter optimization, which facilitates progressive domain adaptation by first learning broad agricultural features before achieving gradual feature specialization on the final target domain. Finally, the model is fine-tuned on a target dataset of 473 images collected from tomato, cucumber, and tobacco greenhouse crops in southern Tunisia. Results demonstrate that this sequential strategy outperforms direct transfer learning in low-data regimes, with recall increasing by 6.7% and Average Precision (AP50-95) by 7.7% compared to the baseline YOLOv8n. The framework enhances the system’s adaptability to background and lighting variations. These findings highlight sequential transfer learning as a practical solution for deploying tiny insect detection models in real-field environments where data are limited and visual conditions are highly variable.</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 sequential transfer learning framework for whitefly detection in low-data agricultural environments

  • 期刊论文

  • 发表日期：

    2026年7月23日

查看全文
  • Ramzi Habtoul,
  • Marwa Zaabi,
  • Nadia Smaoui

摘要

Whiteflies are one of the most common and harmful pests in agriculture, capable of rapid spreading and severe economic losses. In automated whitefly detection, limited annotated data, background variances, and shifting lighting conditions hinder model generalization. The key contribution of this study lies in the implementation of an intermediate agricultural domain adaptation stage prior to final fine-tuning on the target dataset. The proposed sequential transfer learning approach commences with a deep learning You Only Look Once (YOLOv8) model pretrained on the MSCOCO dataset. Next, the model is exposed to the publicly available Cassava dataset (3,000 whitefly images). This enables intermediate-domain adaptation and hyperparameter optimization, which facilitates progressive domain adaptation by first learning broad agricultural features before achieving gradual feature specialization on the final target domain. Finally, the model is fine-tuned on a target dataset of 473 images collected from tomato, cucumber, and tobacco greenhouse crops in southern Tunisia. Results demonstrate that this sequential strategy outperforms direct transfer learning in low-data regimes, with recall increasing by 6.7% and Average Precision (AP50-95) by 7.7% compared to the baseline YOLOv8n. The framework enhances the system’s adaptability to background and lighting variations. These findings highlight sequential transfer learning as a practical solution for deploying tiny insect detection models in real-field environments where data are limited and visual conditions are highly variable.