摘要
及早识别晚疫病是控制马铃薯晚疫病发展的关键,为充分利用高光谱数据波段间特征信息,提高模型对马铃薯晚疫病早期识别的精度,本文提出一种基于双向时间卷积网络(BiTCN)融合自注意力机制(SA)的马铃薯晚疫病早期识别模型(BiTCN-SA)。BiTCN通过正向和反向卷积支路捕捉波段间相关性特征,充分利用前后波段的关联性;自注意力机制动态分配不同波段的重要性权重,提高关键波段对模型分类的贡献度。BiTCN-SA模型将自注意力与BiTCN相融合,在双向上实现局部卷积特征与全局注意力权重的结合,实现双重特征提取,提高模型识别精度。采集3个等级(健康、无症状期、症状初期)的叶片高光谱数据并建模分析,通过对比SVM、RF等机器学习方法和CNN、LSTM、TCN、BiTCN等深度学习方法,以验证本文模型优越性。结果表明,BiTCN-SA模型的收敛速度比单一TCN和BiTCN更快,且模型精度显著提高,比其他机器学习和深度学习方法,具备更强大的特征提取能力,总体准确率达到98%,且对无症状期的病叶识别率达到96%。该方法充分利用高光谱波段间的深层信息,且模型识别率相比于其他机器学习和深度学习方法有大幅提高,为马铃薯晚疫病早期预警和防治提供技术支持。
Early identification of potato late blight is crucial for controlling its development.To fully utilize the inter-band characteristic information of hyperspectral data and improve the accuracy of models in the early identification of potato late blight,this study proposes a potato late blight early identification model(BiTCN-SA)based on a Bidirectional Temporal Convolutional Network(BiTCN)fused with a Self-Attention(SA)mechanism.The BiTCN captures inter-band correlation features through forward and backward convolution branches,and fully exploits the associations between preceding and subsequent bands.The self-attention mechanism dynamically assigns importance weights to different bands,enhancing the contribution of key bands to model classification.The BiTCN-SA model integrates self-attention with BiTCN to achieve a combination of local convolutional features and global attention weights in both directions,realizing dual feature extraction and improving the model's identification accuracy.This study collects hyperspectral potato leaf data from three stages(healthy,asymptomatic,and early symptomatic),conducts modeling and analysis.It verifies the superiority of the proposed model by comparing machine learning methods such as SVM and RF,and deep learning models including CNN,LSTM,TCN,and BiTCN.The results show that the BiTCN-SA model converges faster than standalone TCN and BiTCN models,with significantly improved accuracy.It demonstrates stronger feature extraction capability than other machine learning and deep learning methods,achieving an overall accuracy of 98%and an identification rate of 96%for asymptomatic diseased leaves.This method fully utilizes deep inter-band information from hyperspectral data,and its identification rate shows substantial improvement over other machine learning and deep learning methods,providing technical support for early warning and control of potato late blight.
作者
罗祖升
刘雨琛
王晓丹
张巧杰
LUO Zu-sheng;LIU Yu-chen;WANG Xiao-dan;ZHANG Qiao-jie(College of Automation/Beijing Information Science and Technology University,Beijing 100192,China;College of Plant Protection/China Agricultural University,Beijing 100193,China)
出处
《山东农业大学学报(自然科学版)》
北大核心
2026年第1期56-65,共10页
Journal of Shandong Agricultural University:Natural Science Edition
基金
科技创新2030-“新一代人工智能”重大项目(2021ZD0113603)。
关键词
马铃薯晚疫病
高光谱成像
早期识别
双向时间卷积网络
自注意力机制
特征提取
Potato late blight
hyperspectral imaging
early identification
bidirectional temporal convolutional network
selfattention mechanism
feature extraction
