Scholarlink

<p>Accurate identification of crop phenological stages and plant health is vital for precision agriculture. This study introduces an ensemble deep learning framework for black pepper (<i>Piper nigrum</i>&#xa0;L.), a&#xa0;perennial crop highly sensitive to microclimatic variations in tropical regions like Kasaragod, Kerala. The proposed model integrates four optimized CNNs with a&#xa0;dynamic weighted voting mechanism, enhancing feature extraction while reducing training instability and overfitting common in single-network approaches. Using the KSDBPCDS dataset, comprising 10&#xa0;phenological-health categories captured under real field conditions with standardized mobile photography across canopy levels, diurnal lighting, and seasonal variations, the framework demonstrates robust adaptability. Key innovations include: (1)&#xa0;a&#xa0;multi-model ensemble with enhanced feature extraction, (2)&#xa0;an adaptive dynamic weighting system, and (3)&#xa0;comprehensive preprocessing pipelines for field variability. Experiments achieved 97.95% classification accuracy, surpassing conventional single-model methods, with field simulations confirming robustness under natural light fluctuations. This establishes a&#xa0;benchmark for perennial crop monitoring framework for tropical precision agriculture.</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

Dynamic Ensemble Deep Learning for Robust Phenological Stage and Health Classification in Black Pepper Piper Nigrum L.) Cultivation

  • 期刊论文

  • 发表日期：

    2025年10月2日

查看全文
  • Ratheesh Raju,
  • T. M. Thasleema

摘要

Accurate identification of crop phenological stages and plant health is vital for precision agriculture. This study introduces an ensemble deep learning framework for black pepper (Piper nigrum L.), a perennial crop highly sensitive to microclimatic variations in tropical regions like Kasaragod, Kerala. The proposed model integrates four optimized CNNs with a dynamic weighted voting mechanism, enhancing feature extraction while reducing training instability and overfitting common in single-network approaches. Using the KSDBPCDS dataset, comprising 10 phenological-health categories captured under real field conditions with standardized mobile photography across canopy levels, diurnal lighting, and seasonal variations, the framework demonstrates robust adaptability. Key innovations include: (1) a multi-model ensemble with enhanced feature extraction, (2) an adaptive dynamic weighting system, and (3) comprehensive preprocessing pipelines for field variability. Experiments achieved 97.95% classification accuracy, surpassing conventional single-model methods, with field simulations confirming robustness under natural light fluctuations. This establishes a benchmark for perennial crop monitoring framework for tropical precision agriculture.