Scholarlink

<p>The early and accurate stress detection of plants is key to maximizing productivity, as well as preventing agricultural losses in the long term. Even though deep learning models have been recorded as the most accurate feature of classifying foliar diseases based on visual representation of foliar symptoms, they may not indicate the variation between similarities in symptoms caused by nutrient deficiencies leading, to incorrect diagnoses and treatments. This study proposes the use of a&#xa0;hybrid diagnostic system, which is a&#xa0;combination of an architecture of deep convolutional neural networks (EfficientNet, MobileNet, DenseNet, and ResNet) and a&#xa0;nutrient stress classifier, which is implemented in the form of extreme gradient boosting (XGBoost), trained on 11&#xa0;physiological measurements, including soil pH, soil moisture, nitrogen, phosphorus, and electrochemical signals. The disease ensemble classification is 99.80% correct in five conditions of apple leaf, and the nutrient stress classifier is 99.58% correct in identifying diseases of high, moderate, and healthy stress. The proposed multimodal framework achieved 99.8% accuracy on an unseen test set following stratified 70/15/15 data splitting and five-fold cross-validation, demonstrating strong generalization capability and robustness. It is proposed to adopt a&#xa0;decision-level fusion approach to make interpretation of predictions and make context-based mitigation between fungicidal spraying, nutrient supplementation, and no action. To make these more transparent and interpretable, we visualize model attention on leaf images with gradient-weighted class activation mapping (Grad-CAM), visualize feature contribution in nutrient stress predictions with SHapley Additive exPlanations (SHAP), and project the learned feature distributions with t‑distributed stochastic neighbor embedding (t-SNE). All the results are validated on real test data and presented in a&#xa0;format that is easy to understand, which represents a&#xa0;huge step towards the intelligent multimodal measurement of plant health.</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 Vision-Sensor Framework for Apple Leaf Disease and Health Diagnosis Using Deep Ensemble Learning and Nutrient Profiling

  • 期刊论文

  • 发表日期：

    2026年3月25日

查看全文
  • Saimul Bashir,
  • Adil Bashir

摘要

The early and accurate stress detection of plants is key to maximizing productivity, as well as preventing agricultural losses in the long term. Even though deep learning models have been recorded as the most accurate feature of classifying foliar diseases based on visual representation of foliar symptoms, they may not indicate the variation between similarities in symptoms caused by nutrient deficiencies leading, to incorrect diagnoses and treatments. This study proposes the use of a hybrid diagnostic system, which is a combination of an architecture of deep convolutional neural networks (EfficientNet, MobileNet, DenseNet, and ResNet) and a nutrient stress classifier, which is implemented in the form of extreme gradient boosting (XGBoost), trained on 11 physiological measurements, including soil pH, soil moisture, nitrogen, phosphorus, and electrochemical signals. The disease ensemble classification is 99.80% correct in five conditions of apple leaf, and the nutrient stress classifier is 99.58% correct in identifying diseases of high, moderate, and healthy stress. The proposed multimodal framework achieved 99.8% accuracy on an unseen test set following stratified 70/15/15 data splitting and five-fold cross-validation, demonstrating strong generalization capability and robustness. It is proposed to adopt a decision-level fusion approach to make interpretation of predictions and make context-based mitigation between fungicidal spraying, nutrient supplementation, and no action. To make these more transparent and interpretable, we visualize model attention on leaf images with gradient-weighted class activation mapping (Grad-CAM), visualize feature contribution in nutrient stress predictions with SHapley Additive exPlanations (SHAP), and project the learned feature distributions with t‑distributed stochastic neighbor embedding (t-SNE). All the results are validated on real test data and presented in a format that is easy to understand, which represents a huge step towards the intelligent multimodal measurement of plant health.