Scholarlink

<p>Plant diseases remain a major constraint on global agricultural productivity, causing substantial economic losses and threatening food security. Although recent deep learning models have achieved remarkable classification performance, many remain computationally expensive, insufficiently interpretable, and incapable of accurately characterizing stochastic disease transition zones where healthy and infected tissues coexist. Furthermore, severity estimation, which is essential for practical agronomic decision-making, is often overlooked in existing edge-intelligence frameworks. To address these challenges, this study proposes a Bayesian-Optimized Hybrid Fuzzy–Probabilistic Neural Network (Fuzzy–PNN) framework for ultra-lightweight severity-aware disease monitoring in precision agricultural Internet of Things (IoT) environments. The proposed architecture integrates CIE-Lab and HSV color transformation, K-means clustering, Mask R-CNN-assisted lesion localization, Gray-Level Co-occurrence Matrix (GLCM)-based texture characterization, Mamdani fuzzy inference, Bayesian optimization, and probabilistic neural classification within a unified decision framework. Disease severity is quantified using a Weighted Severity Index (WSI), enabling continuous assessment of pathological progression rather than discrete disease categorization alone. Experimental evaluation was conducted using 4500 PlantVillage images comprising potato, tomato, and pepper leaf samples under a balanced training protocol. The proposed framework achieved an overall classification accuracy of 98.33%, ROC–AUC of 0.985, F1-score of 0.98, Matthews Correlation Coefficient of 0.98, and Cohen’s κ of 0.97. Statistical benchmarking using Friedman ranking, Wilcoxon signed-rank testing, and Nemenyi post-hoc analysis demonstrated competitive superiority over contemporary edge-intelligence architectures while requiring only 1.22&#xa0;MB memory and 0.58&#xa0;ms inference latency. These findings indicate that the proposed Bayesian-optimized Fuzzy–PNN architecture provides an interpretable, computationally efficient, and deployment-ready solution for next-generation precision agricultural monitoring systems.</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 Bayesian-optimized hybrid fuzzy–PNN framework for ultra-lightweight severity-aware monitoring of stochastic plant disease transition zones in precision agricultural IoT

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月28日

查看全文
  • M. R. K. Pathan,
  • Saiful Islam,
  • Abdul Motaleb,
  • Sajia Afrin

摘要

Plant diseases remain a major constraint on global agricultural productivity, causing substantial economic losses and threatening food security. Although recent deep learning models have achieved remarkable classification performance, many remain computationally expensive, insufficiently interpretable, and incapable of accurately characterizing stochastic disease transition zones where healthy and infected tissues coexist. Furthermore, severity estimation, which is essential for practical agronomic decision-making, is often overlooked in existing edge-intelligence frameworks. To address these challenges, this study proposes a Bayesian-Optimized Hybrid Fuzzy–Probabilistic Neural Network (Fuzzy–PNN) framework for ultra-lightweight severity-aware disease monitoring in precision agricultural Internet of Things (IoT) environments. The proposed architecture integrates CIE-Lab and HSV color transformation, K-means clustering, Mask R-CNN-assisted lesion localization, Gray-Level Co-occurrence Matrix (GLCM)-based texture characterization, Mamdani fuzzy inference, Bayesian optimization, and probabilistic neural classification within a unified decision framework. Disease severity is quantified using a Weighted Severity Index (WSI), enabling continuous assessment of pathological progression rather than discrete disease categorization alone. Experimental evaluation was conducted using 4500 PlantVillage images comprising potato, tomato, and pepper leaf samples under a balanced training protocol. The proposed framework achieved an overall classification accuracy of 98.33%, ROC–AUC of 0.985, F1-score of 0.98, Matthews Correlation Coefficient of 0.98, and Cohen’s κ of 0.97. Statistical benchmarking using Friedman ranking, Wilcoxon signed-rank testing, and Nemenyi post-hoc analysis demonstrated competitive superiority over contemporary edge-intelligence architectures while requiring only 1.22 MB memory and 0.58 ms inference latency. These findings indicate that the proposed Bayesian-optimized Fuzzy–PNN architecture provides an interpretable, computationally efficient, and deployment-ready solution for next-generation precision agricultural monitoring systems.