Scholarlink

<p>Yellow rust and brown rust pose severe threats to wheat crop production and agricultural sustainability. Leveraging technologies such as remote sensing, drones, and sensor networks provide timely and accurate insights into disease outbreaks. Existing UAV-based disease detection frameworks often struggle to balance model accuracy with computational efficiency, limiting their applicability in real-time field conditions. This study proposes a novel hybrid lightweight architecture specifically optimized for deployment on unmanned aerial vehicles (UAVs) and remote sensors for real-time wheat disease identification. The proposed model integrated MobileNetV2 architecture with a single 3 × 3 Inception(V4-inspired) Model. The lightweight MobileNetV2 extracts relevant features from the drone captured image, while Inception(V4-inspired) Model excels at extracting complex and multi-scale features. By using just one Inception(V4-inspired) Model, the overall model size and computational cost remain relatively low. A total of 362 real UAV-captured images were used as the primary dataset and later amplified to 3,727 images via augmentation techniques to improve balance and model generalization. Comparative analysis is conducted with five other advanced lightweight models, including InceptionV3, NASNetMobile, MobileNet, MobileNetV2, and DenseNet121. The proposed lightweight model demonstrates high accuracy, reaching 92.08% after 20 training epochs and improving to 93.75% after 30 epochs. This is achieved with a compact design with only 2,517,044 training parameters. Training efficiency is further enhanced with an average training time of 82.95&#xa0;s per epoch, which reduces to 71.73&#xa0;s after 30 epochs. The proposed model surpasses other models by exhibiting superior performance, including higher accuracy and reduced computational complexity. Beyond its technical contribution, this work highlights the potential of integrating lightweight deep learning architectures with UAV platforms to advance precision agriculture and sustainable crop monitoring.</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 hybrid lightweight model for real-time wheat disease identification using Unmanned Aerial Vehicles (UAV) imaging

  • 期刊论文

  • 发表日期：

    2026年2月26日

查看全文
  • Rajeev Singh,
  • Akash Arya,
  • km Neha

摘要

Yellow rust and brown rust pose severe threats to wheat crop production and agricultural sustainability. Leveraging technologies such as remote sensing, drones, and sensor networks provide timely and accurate insights into disease outbreaks. Existing UAV-based disease detection frameworks often struggle to balance model accuracy with computational efficiency, limiting their applicability in real-time field conditions. This study proposes a novel hybrid lightweight architecture specifically optimized for deployment on unmanned aerial vehicles (UAVs) and remote sensors for real-time wheat disease identification. The proposed model integrated MobileNetV2 architecture with a single 3 × 3 Inception(V4-inspired) Model. The lightweight MobileNetV2 extracts relevant features from the drone captured image, while Inception(V4-inspired) Model excels at extracting complex and multi-scale features. By using just one Inception(V4-inspired) Model, the overall model size and computational cost remain relatively low. A total of 362 real UAV-captured images were used as the primary dataset and later amplified to 3,727 images via augmentation techniques to improve balance and model generalization. Comparative analysis is conducted with five other advanced lightweight models, including InceptionV3, NASNetMobile, MobileNet, MobileNetV2, and DenseNet121. The proposed lightweight model demonstrates high accuracy, reaching 92.08% after 20 training epochs and improving to 93.75% after 30 epochs. This is achieved with a compact design with only 2,517,044 training parameters. Training efficiency is further enhanced with an average training time of 82.95 s per epoch, which reduces to 71.73 s after 30 epochs. The proposed model surpasses other models by exhibiting superior performance, including higher accuracy and reduced computational complexity. Beyond its technical contribution, this work highlights the potential of integrating lightweight deep learning architectures with UAV platforms to advance precision agriculture and sustainable crop monitoring.