Scholarlink

<p>National income receives major contributions from the vital&#xa0;agriculture sector. The expanding worldwide population requires continual innovation that boosts agricultural production because of increasing demands. This study investigates how Deep Learning (DL) models such as the ResNet can&#xa0;serve smart agriculture purposes. The paper provides an extensive evaluation of the top Convolutional Neural Networks (CNN) models with a particular focus on ResNet for smart agricultural applications. This research work&#xa0;investigates customized ResNet implementations along with their functionality and performance effects for agriculture purposes through analysis of strengths and weaknesses and constraint evaluation for multiple agricultural activities. Our research sets itself apart from other studies, which mainly concentrated on singular agricultural applications, because it offers a unique assessment of different ResNet-based models jointly with their implementations across multiple smart agriculture fields. Deep Learning (DL) technology advances show that deeper neural networks can help achieve better results at the expense of greater network complexity. The depth of neural networks causes them to develop the vanishing gradient problem that results in performance loss and decreased efficiency. The ResNet architecture delivers high performance quality through its ability to maintain unaltered network depth, therefore presenting itself as an appealing solution for smart agriculture applications. Research findings demonstrate that transformer-based models (ResViT-Rice, T-RNet) achieve superior performance compared&#xa0;to standard CNNs, because they effectively track dependencies across extensive backgrounds that are of interest for complex background analysis. The combination of transfer learning and feature fusion capabilities in TL-ResUNet makes it better than DeepLabv3 + &#xa0; in&#xa0;performing land classification tasks. The class imbalance handling capability of T-RNet makes it most beneficial for diagnosing rare diseases, effectively. The bilinear pooling technique of DIR-BiRN provides precise detection of subtle diseases but requires significant computational power. The ResNet-50 Parallel Channel Spatial Attention (PCSA)&#xa0;operative model provides efficient pest detection; however, its vulnerability prevents it from handling datasets with high variability. GAN-based data augmentation confirmed effective in beef quality identification through improved classification results; however, its utility is restricted by the scarcity of realistic diversity within synthetic samples. Among the checked publications, the highest accuracy is achieved with the ResNet-18 model, at a level o 99% using segmentation techniques. The second highest accuracy reported is 98.17% using the ResNet-50 PSCA model. The third highest accuracy obtained is 97.84% using the ResViT-Rice model. The fourth highest accuracy informed is 96.76% using the DIR-BiRN model, whereas the fifth highest accuracy reported is 95.83% using the pre-trained ResNet model. The results emphasize the effectiveness of the&#xa0;ResNet&#xa0;model in advancing smart agriculture, demonstrating its potential as a robust architecture for agricultural applications.</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

Comprehensive survey on residual neural networks improvements in agricultural applications

  • 期刊论文

  • 发表日期：

    2025年11月22日

查看全文
  • Elhossiny Ibrahim,
  • Medhat Hamdy,
  • Adel S. El-Fishawy,
  • Sami A Eldolil,
  • Heba El-Hoseny

摘要

National income receives major contributions from the vital agriculture sector. The expanding worldwide population requires continual innovation that boosts agricultural production because of increasing demands. This study investigates how Deep Learning (DL) models such as the ResNet can serve smart agriculture purposes. The paper provides an extensive evaluation of the top Convolutional Neural Networks (CNN) models with a particular focus on ResNet for smart agricultural applications. This research work investigates customized ResNet implementations along with their functionality and performance effects for agriculture purposes through analysis of strengths and weaknesses and constraint evaluation for multiple agricultural activities. Our research sets itself apart from other studies, which mainly concentrated on singular agricultural applications, because it offers a unique assessment of different ResNet-based models jointly with their implementations across multiple smart agriculture fields. Deep Learning (DL) technology advances show that deeper neural networks can help achieve better results at the expense of greater network complexity. The depth of neural networks causes them to develop the vanishing gradient problem that results in performance loss and decreased efficiency. The ResNet architecture delivers high performance quality through its ability to maintain unaltered network depth, therefore presenting itself as an appealing solution for smart agriculture applications. Research findings demonstrate that transformer-based models (ResViT-Rice, T-RNet) achieve superior performance compared to standard CNNs, because they effectively track dependencies across extensive backgrounds that are of interest for complex background analysis. The combination of transfer learning and feature fusion capabilities in TL-ResUNet makes it better than DeepLabv3 +   in performing land classification tasks. The class imbalance handling capability of T-RNet makes it most beneficial for diagnosing rare diseases, effectively. The bilinear pooling technique of DIR-BiRN provides precise detection of subtle diseases but requires significant computational power. The ResNet-50 Parallel Channel Spatial Attention (PCSA) operative model provides efficient pest detection; however, its vulnerability prevents it from handling datasets with high variability. GAN-based data augmentation confirmed effective in beef quality identification through improved classification results; however, its utility is restricted by the scarcity of realistic diversity within synthetic samples. Among the checked publications, the highest accuracy is achieved with the ResNet-18 model, at a level o 99% using segmentation techniques. The second highest accuracy reported is 98.17% using the ResNet-50 PSCA model. The third highest accuracy obtained is 97.84% using the ResViT-Rice model. The fourth highest accuracy informed is 96.76% using the DIR-BiRN model, whereas the fifth highest accuracy reported is 95.83% using the pre-trained ResNet model. The results emphasize the effectiveness of the ResNet model in advancing smart agriculture, demonstrating its potential as a robust architecture for agricultural applications.