基于密集卷积网络的甘薯黑斑病害程度分类  

Classification of sweet potato black spot disease severity using dense convolutional networks

作　　者：张帅杰 陈思思 戴丹[1] 李艳宏 梁子乐 罗煦钦 霍富龙 胡彦蓉[1] ZHANG Shuaijie;CHEN Sisi;DAI Dan;LI Yanhong;LIANG Zile;LUO Xuqin;HUO Fulong;HU Yanrong(School of Mathematics and Computer Science,Zhejiang Agriculture and Forestry University,Hangzhou 311300,China;Hangzhou Lin'an District Agricultural and Forestry Technology Extension Center,Hangzhou 311300,China;Hangzhou Lin'an District Agricultural and Rural Information Service Center,Hangzhou 311300,China)

机构地区：[1]浙江农林大学数学与计算机科学学院,杭州311300 [2]杭州市临安区农林技术推广中心,杭州311300 [3]杭州市临安区农业农村信息服务中心,杭州311300

出　　处：《中国农业大学学报》2025年第4期51-66,共16页Journal of China Agricultural University

基　　金：国家青年科学基金项目(32301585,42001354);浙江省重大科技专项重点农业项目(2015C02047)。

摘　　要：针对长喙壳菌侵染导致的甘薯块根黑斑病严重影响薯块品质和加工产品食用安全问题,以人工培养环境下获得的甘薯黑斑病图像为研究对象,选取总酚质量分数为内在指标,黑斑直径为外观指标,利用机器学习探索两者之间的相关性,并据此进行病害等级划分。对比分析了ResNet152、VGG19、GoogleNet、EfficientNetB7、MobileNetV2和DenseNet201模型的检测效果后,提出一种融合空间注意力机制(SA)、高效通道注意力机制(ECA)和DenseNet201网络的算法,用于甘薯黑斑病害识别研究。结果表明:1)黑斑直径与总酚质量分数的皮尔逊相关系数为0.93,呈显著正相关(P<0.01),表明甘薯外部病害特征可有效反映内部品质变化;2)在不同病害等级的甘薯黑斑病害识别中,DenseNet201模型的准确率最高,达到83.93%;3)消融实验表明,引入SA和ECA机制后构建的DenseNet201-SA-ECA模型在测试集上的分类准确率达96.8%,较原DenseNet201模型提高了12.87%,显著提升了对黑斑病害的精准识别。本研究结合深度学习与生化指标所构建的DenseNet201-SA-ECA模型在甘薯黑斑病害识别方面的性能显著优于其他卷积神经网络,可实现对甘薯黑斑病的准确识别。To address the problem of sweet potato black spot disease caused by Ceratocystis fimbriata,which severely impacts tuber quality and the safety of processed products,this study utilized images of sweet potato black spot disease obtained from controlled cultivation environments.Total phenolic content was selected as an internal quality indicator,while black spot diameter was used as an external appearance indicator.Machine learning techniques were applied to explore the correlation between these two factors,and based on the results,disease severity grading was performed.After comparing the detection performance of ResNet152,VGG19,GoogleNet,EfficientNetB7,MobileNetV2,and DenseNet201 models,a novel algorithm that integrates the Spatial Attention(SA)mechanism,Efficient Channel Attention(ECA)mechanism,and the DenseNet201 network was proposed for the identification of sweet potato black spot disease.The results showed that:1)The Pearson correlation coefficient between black spot diameter and total phenolic content was 0.93,indicating a strong positive correlation(P<0.01),which suggested that external disease symptoms can effectively reflect internal quality changes;2)Among the models used to identify different disease grades of sweet potato black spot disease,DenseNet201 achieved the highest accuracy of 83.93%;3)The results of Ablation experiments demonstrated that the DenseNet201-SA-ECA model,which incorporated both SA and ECA mechanisms,achieved a classification accuracy of 96.8%on the test set,representing a 12.87%improvement over the original DenseNet201 model and significantly enhancing the precision of black spot disease identification.Combining biochemical indicators with deep learning,this study shows that the DenseNet201-SA-ECA model outperforms other convolutional neural networks in identifying sweet potato black spot disease and can reliably recognize the disease with high accuracy.

关 键 词：甘薯 病害分类 深度学习 总酚质量分数 卷积网络模型 

分 类 号：S531[农业科学—作物学]