Scholarlink

<p>Black gram (Vigna mungo L.) is an important protein rich pulse crop, which is widely grown in south Asia. Foliar diseases, however, like leaf crinkle, anthracnose, yellow mosaic and powdery mildew have a significant effect on its productivity. For minority crops there is often a lack of large, annotated datasets for existing methods of disease detection. Moreover, the majority of deep learning models are not very interpretable, and it is hard to audit their decisions in low-resource field settings. In this study, a new interpretable Vein-Lesion Contrastive Morphology (VLCM) framework for black gram leaf disease detection in limited-data condition is proposed to overcome the above limitations. The model proposed is able to extract three complementary streams of visual evidence: Lesion intensity analysis, vein disruption morphometry and texture degradation profiling. The proposed framework extracts three complementary evidence streams comprising lesion intensity analysis, vein disruption morphometry, and texture degradation profiling. These evidence streams are fused through a Symbolic Disease Score (SDS), an interpretable intermediate representation that combines their weighted contributions into a single scalar measure of disease severity while supporting the final ensemble classification. To achieve better discrimination between disease classes, a feature weighting method which combines joint Bhattacharyya separability and Shannon entropy is used. Finally, the so called ensemble classifier is formed by support vector machine and gradient boosting classifiers to classify five-classes of black gram leaves. The experimental evaluation was carried out on BPLD dataset. The proposed framework VLCM was found to attain the mean accuracy of 96.8% and macro F1-score of 96.4%. The results were statistically validated with Wilcoxon signed-rank test, which was found to be significant as compared to existing methods at 0.001 level of significance. The model also had a very small size of 102 milliseconds per image on a commodity CPU, suitable for deployment on the field without the use of special hardware. The outcomes reveal that the proposed VLCM framework has the ability to offer a precise, interpretable and computational efficient solution to recognize the black gram leaf disease in resource-restricted agricultural setting.</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

An Interpretable Vein-Lesion Contrastive Morphology Framework for Black Gram Leaf Disease Recognition

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年8月1日

查看全文
  • S. Ponmythili,
  • D. Janaki Sathya

摘要

Black gram (Vigna mungo L.) is an important protein rich pulse crop, which is widely grown in south Asia. Foliar diseases, however, like leaf crinkle, anthracnose, yellow mosaic and powdery mildew have a significant effect on its productivity. For minority crops there is often a lack of large, annotated datasets for existing methods of disease detection. Moreover, the majority of deep learning models are not very interpretable, and it is hard to audit their decisions in low-resource field settings. In this study, a new interpretable Vein-Lesion Contrastive Morphology (VLCM) framework for black gram leaf disease detection in limited-data condition is proposed to overcome the above limitations. The model proposed is able to extract three complementary streams of visual evidence: Lesion intensity analysis, vein disruption morphometry and texture degradation profiling. The proposed framework extracts three complementary evidence streams comprising lesion intensity analysis, vein disruption morphometry, and texture degradation profiling. These evidence streams are fused through a Symbolic Disease Score (SDS), an interpretable intermediate representation that combines their weighted contributions into a single scalar measure of disease severity while supporting the final ensemble classification. To achieve better discrimination between disease classes, a feature weighting method which combines joint Bhattacharyya separability and Shannon entropy is used. Finally, the so called ensemble classifier is formed by support vector machine and gradient boosting classifiers to classify five-classes of black gram leaves. The experimental evaluation was carried out on BPLD dataset. The proposed framework VLCM was found to attain the mean accuracy of 96.8% and macro F1-score of 96.4%. The results were statistically validated with Wilcoxon signed-rank test, which was found to be significant as compared to existing methods at 0.001 level of significance. The model also had a very small size of 102 milliseconds per image on a commodity CPU, suitable for deployment on the field without the use of special hardware. The outcomes reveal that the proposed VLCM framework has the ability to offer a precise, interpretable and computational efficient solution to recognize the black gram leaf disease in resource-restricted agricultural setting.