Scholarlink

<p>Wheat (<i>Triticum aestivum</i>) is an important cereal crop and a key source of food and nutrition in the world. Its growth and productivity are heavily influenced by nitrogen, an essential macronutrient vital for plant metabolism and yield. Accurate estimation of nitrogen content is crucial in wheat crops to improve nitrogen fertilizer use efficiency, improve crop productivity, and minimize environmental impact. However, existing nitrogen estimation methods often overlook the influence of environmental variables, labour-intensive, destructive, and not suitable for real-time monitoring. The primary objective of this study is to develop and validate a deep learning framework that integrates multi-source data, specifically RGB leaf images and environmental parameters (air temperature, humidity, soil temperature) for accurate nitrogen content estimation in wheat. The framework combines visual features extracted from leaf images with contextual environmental data to capture both phenotypic and physiological factors influencing nitrogen status. A total of 2232 wheat leaf images collected under controlled nitrogen treatments were pre-processed using two approaches: (i) a U<sup>2</sup>-Net based background removal and normalisation pipeline, and (ii) a more complex multi-stage enhancement process involving thresholding, edge detection, and contrast adjustments. Leaf nitrogen content was determined using the standard Kjeldahl laboratory method, providing ground-truth values for model training and validation. Transfer learning models, Xception and DenseNet121, were trained under four experimental configurations that combined image and weather inputs. The DenseNet121 model with U<sup>2</sup>-Net pre-processing and environmental variables (air temperature, humidity, and soil temperature) achieved the best performance with an RMSE of 0.693, outperforming both Xception (0.713) and complex pre-processing approaches (0.835). K-Fold cross-validation confirmed robust within-season performance (RMSE = 0.286 for 2022–23; 0.470 for 2023–24), while cross-season testing highlighted reduced generalisation due to environmental variability. The results demonstrate that integrating environmental covariates with deep image features substantially enhances the accuracy and scalability of nitrogen estimation. The proposed DenseNet121 + U<sup>2</sup>-Net + weather configuration provides a cost-effective and generalizable framework for real-time nitrogen monitoring, supporting more precise and sustainable fertilizer management in 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

Integration of environmental variables with deep learning models for nitrogen content estimation in wheat leaves

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月2日

查看全文
  • N. Bharath Kumar,
  • Krishna Kumar Chaturvedi,
  • Rajeev Ranjan Kumar,
  • S. C. Tripathi,
  • R. P. Meena,
  • Anu Sharma,
  • Shashi Bhushan Lal,
  • Mohammad Samir Farooqi,
  • Sanjeev Kumar,
  • Rajeev Ranjan

摘要

Wheat (Triticum aestivum) is an important cereal crop and a key source of food and nutrition in the world. Its growth and productivity are heavily influenced by nitrogen, an essential macronutrient vital for plant metabolism and yield. Accurate estimation of nitrogen content is crucial in wheat crops to improve nitrogen fertilizer use efficiency, improve crop productivity, and minimize environmental impact. However, existing nitrogen estimation methods often overlook the influence of environmental variables, labour-intensive, destructive, and not suitable for real-time monitoring. The primary objective of this study is to develop and validate a deep learning framework that integrates multi-source data, specifically RGB leaf images and environmental parameters (air temperature, humidity, soil temperature) for accurate nitrogen content estimation in wheat. The framework combines visual features extracted from leaf images with contextual environmental data to capture both phenotypic and physiological factors influencing nitrogen status. A total of 2232 wheat leaf images collected under controlled nitrogen treatments were pre-processed using two approaches: (i) a U2-Net based background removal and normalisation pipeline, and (ii) a more complex multi-stage enhancement process involving thresholding, edge detection, and contrast adjustments. Leaf nitrogen content was determined using the standard Kjeldahl laboratory method, providing ground-truth values for model training and validation. Transfer learning models, Xception and DenseNet121, were trained under four experimental configurations that combined image and weather inputs. The DenseNet121 model with U2-Net pre-processing and environmental variables (air temperature, humidity, and soil temperature) achieved the best performance with an RMSE of 0.693, outperforming both Xception (0.713) and complex pre-processing approaches (0.835). K-Fold cross-validation confirmed robust within-season performance (RMSE = 0.286 for 2022–23; 0.470 for 2023–24), while cross-season testing highlighted reduced generalisation due to environmental variability. The results demonstrate that integrating environmental covariates with deep image features substantially enhances the accuracy and scalability of nitrogen estimation. The proposed DenseNet121 + U2-Net + weather configuration provides a cost-effective and generalizable framework for real-time nitrogen monitoring, supporting more precise and sustainable fertilizer management in precision agriculture.