基于光谱优化和Transformer的香蕉叶斑病早期检测方法

Early Detection Method for Banana Leaf Spot Disease Based on Spectral Optimization and Transformer

摘要:

香蕉叶斑病是香蕉生产危害叶片最常见的真菌性病害之一,其潜伏期诊断对于实现早期的精准防控与优化农药施用策略具有重要意义。本文提出一种结合光谱优化方法和Transformer的模型,构建了香蕉叶斑病早期诊断模型。模型基于健康、潜伏期和发病的香蕉叶斑病苗期叶片高光谱图像为研究对象,提取其感兴趣区域内平均光谱,通过对光谱预处理方法和特征波段选择方法进行研究,排列组合得到30种光谱优化处理方法,确定最佳光谱优化处理方法是卷积平滑法（SG）结合一阶导数（D1）预处理,再使用竞争自适应重加权算法（CARS）进行波段选择。基于优化光谱数据构建的Transformer模型总体准确率达91.78%,较未优化光谱模型提升12.21个百分点。试验结果表明,本文提出的SG-D1-CARS光谱优化处理方法能够有效地减少光谱数据间的影响,消除冗余信息,优化光谱谱带,为实现香蕉叶斑病早期检测提供了理论和技术支持。

Abstract:

A rapid early diagnosis model for banana leaf spot disease was constructed and evaluated by combining spectral optimization methods with a Transformer model. The investigation was motivated by the fact that banana leaf spot disease is among the most common fungal diseases that damage banana leaves. It is widely recognized that early diagnosis,achieved before symptom manifestation,is crucial for enabling precise early intervention and optimized pesticide use. The hyperspectral images of healthy,diseased,and asymptomatic samples were collected,and the average spectral data of region of interest between 400 nm and 900 nm bands was extracted. In response to the problem of mutual interference in hyperspectral data of banana leaf spot disease with different severity levels,as well as the decrease in prediction accuracy caused by the large number of bands,large data volume,and complex spectral bands,spectral preprocessing methods and feature band selection methods were studied. Four preprocessing algorithms,Savitzky-Golay convolution smoothing (SG),standard normal transformation (SNV),first derivative (D1),and multiple scattering correction (MSC),and their combined model effects were compared and analyzed. Principal component analysis (PCA),continuous projection algorithm (SPA),and competitive adaptive reweighting algorithm (CARS) were used for feature wavelength extraction. By combining these methods,a total of 30 spectral processing methods were obtained,which were then combined with the Transformer model with center loss added to establish optimized discriminant models. Based on the average accuracy of the established discriminant models,the optimal spectral processing method suitable for early detection of banana leaf spot disease was finally selected. The research results showed that the classification accuracy of optimized spectral data using 22 spectral processing methods was significantly improved compared with the overall classification accuracy of 79.57% for the original data. The optimal spectral processing method was SG-D1-CARS,combined with the confusion matrix. The Transformer model constructed based on optimized spectral data achieved an overall accuracy of 91.78%,an improvement of 12.21 percentage points compared with that of the unoptimized spectral model. The experimental results indicated that the proposed method can effectively detect banana leaf spot disease in the early stage. By combining various preprocessing and feature extraction methods based on optimization objectives,the modeling effect can be effectively improved,providing theoretical and technical support for non-destructive,accurate,and intelligent detection of crop diseases.