Scholarlink

The increasing prevalence of fungal infections in banana cultivation specifically Sigatoka, Cordana, and Pestalotiopsis poses a serious threat to crop yield and global agricultural sustainability. This study aims to develop an accurate, interpretable classification model for early-stage detection of banana leaf diseases using image-based deep learning approaches. The BananaLSD dataset, containing annotated leaf images, was utilized to classify samples into four categories: Healthy, Sigatoka, Cordana, and Pestalotiopsis. Four state-of-the-art Convolutional Neural Network (CNN) architectures—EfficientNet-B0, ResNet-50, MobileNet-V2, and SqueezeNet were trained and evaluated using standard metrics. EfficientNet-B0 achieved the highest performance with an accuracy, precision, and recall of 98.41%, significantly outperforming the other models. To address the black-box nature of CNNs and enhance model interpretability, LIME (Local Interpretable Model-agnostic Explanations) and SHAP (SHapley Additive exPlanations) were employed to visualize salient features influencing predictions. The findings confirm the viability of EfficientNet-B0 for high-accuracy classification of banana leaf diseases and demonstrate the utility of XAI frameworks in making AI-driven agricultural diagnostics more transparent and trustworthy.

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

An Explainable Deep Learning Framework for Banana Leaf Disease Detection and Analysis

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • M. Amritha,
  • Lidiya Lilly Thampi,
  • Berin Pathrose

摘要

The increasing prevalence of fungal infections in banana cultivation specifically Sigatoka, Cordana, and Pestalotiopsis poses a serious threat to crop yield and global agricultural sustainability. This study aims to develop an accurate, interpretable classification model for early-stage detection of banana leaf diseases using image-based deep learning approaches. The BananaLSD dataset, containing annotated leaf images, was utilized to classify samples into four categories: Healthy, Sigatoka, Cordana, and Pestalotiopsis. Four state-of-the-art Convolutional Neural Network (CNN) architectures—EfficientNet-B0, ResNet-50, MobileNet-V2, and SqueezeNet were trained and evaluated using standard metrics. EfficientNet-B0 achieved the highest performance with an accuracy, precision, and recall of 98.41%, significantly outperforming the other models. To address the black-box nature of CNNs and enhance model interpretability, LIME (Local Interpretable Model-agnostic Explanations) and SHAP (SHapley Additive exPlanations) were employed to visualize salient features influencing predictions. The findings confirm the viability of EfficientNet-B0 for high-accuracy classification of banana leaf diseases and demonstrate the utility of XAI frameworks in making AI-driven agricultural diagnostics more transparent and trustworthy.