Scholarlink

<p>AI and data analytics are underpinning a farming technology innovation and assisting in enhanced practices of farming practices. The infestation of pests is one of the most essential predicaments in this field, heavily influencing the output of crops, ecology, and the economy. In this research, HDL-Net is proposed as an early pest detection system with the use of pest sound analytics. The proposed system utilized the pest audio sounds and performed the proactive audio pre-processing schema with the help of efficient audio pre-processing methods such as GFCC, FFT, DFT, Kaiser Window, Blackman-Harris Window, and FPF, and selected representative features to enable comprehensive processing. The 3,800 audio samples of 38 pest species were utilized to train, validate, and test the HDL-Net framework. The HDL-Net integrates the ResNet-50, DCN-Net, AGF-Net, and ASFP-Net modules in order to allow accurate detection of the pests and to exploit the advantages of the different modules in a complementary way. ResNet-50 is used as the baseline of the HDL-Net, where the spectral characteristics can be learned into hierarchical and multi-scale interpretations. The ASFP-Net adapts spatial features to key spatial pooling, AGF-Net, and DCN-Net process in multiple scales, allowing for increased detection by better extracting discriminative and informative features. Consequently, the HDL-Net can provide accuracy up to the classification and early detection in various agricultural conditions of agriculture. The results of the experimental findings demonstrate better performance, which led to an accountability of 99.77% (99.78% by 5-Fold Cross-Validation) accuracy, 99.34% specificity, 99.90% sensitivity, 99.81% precision, 99.90% recall, and 99.85% F1-score, compared to the benchmark models, including VGG-16, DenseNet, DCNN, ANN, YOLOv5, KNN, ResNet-50, and FRCNN. The contribution of research is the development of an automated and scalable framework, which allows detecting the presence of pests and decent potential in the industrial agricultural practice as a productivity and sustainability-enhancing method.</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

HDL-Net: Hybrid deep learning and IoT Network-based system for pest detection using pest sound analytics

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年10月8日

查看全文
  • Md. Akkas Ali,
  • Md. Shohel Sayeed,
  • Siti Fatimah Abdul Razak

摘要

AI and data analytics are underpinning a farming technology innovation and assisting in enhanced practices of farming practices. The infestation of pests is one of the most essential predicaments in this field, heavily influencing the output of crops, ecology, and the economy. In this research, HDL-Net is proposed as an early pest detection system with the use of pest sound analytics. The proposed system utilized the pest audio sounds and performed the proactive audio pre-processing schema with the help of efficient audio pre-processing methods such as GFCC, FFT, DFT, Kaiser Window, Blackman-Harris Window, and FPF, and selected representative features to enable comprehensive processing. The 3,800 audio samples of 38 pest species were utilized to train, validate, and test the HDL-Net framework. The HDL-Net integrates the ResNet-50, DCN-Net, AGF-Net, and ASFP-Net modules in order to allow accurate detection of the pests and to exploit the advantages of the different modules in a complementary way. ResNet-50 is used as the baseline of the HDL-Net, where the spectral characteristics can be learned into hierarchical and multi-scale interpretations. The ASFP-Net adapts spatial features to key spatial pooling, AGF-Net, and DCN-Net process in multiple scales, allowing for increased detection by better extracting discriminative and informative features. Consequently, the HDL-Net can provide accuracy up to the classification and early detection in various agricultural conditions of agriculture. The results of the experimental findings demonstrate better performance, which led to an accountability of 99.77% (99.78% by 5-Fold Cross-Validation) accuracy, 99.34% specificity, 99.90% sensitivity, 99.81% precision, 99.90% recall, and 99.85% F1-score, compared to the benchmark models, including VGG-16, DenseNet, DCNN, ANN, YOLOv5, KNN, ResNet-50, and FRCNN. The contribution of research is the development of an automated and scalable framework, which allows detecting the presence of pests and decent potential in the industrial agricultural practice as a productivity and sustainability-enhancing method.