基于高光谱与叶绿素融合的梨树叶片黑斑病分级检测

Classification Detection of Pear Leaf Black Spot Based on Hyperspectral and Chlorophyll Fusion

摘要:

梨树叶片黑斑病是一种由链格孢属真菌引起的流行性传染病，严重影响梨树的生长发育和果实产量。定性分析梨树叶片黑斑病的感染程度，并实现梨树叶片黑斑病的精准施药，对减少经济损失和实现农药减量使用具有重要意义。本研究基于高光谱成像技术，以不同类型梨树叶片（健康、轻度、中度和重度）为研究对象，通过光谱、图像和叶绿素特征融合，构建梨树叶片黑斑病病害分级检测模型，实现梨树叶片黑斑病高精度分类和分级检测。利用高光谱成像系统获取不同类型梨树叶片高光谱数据，通过SG平滑方法对光谱数据进行预处理，再利用连续投影算法（SPA）和竞争性自适应重加权算法（CARS）提取特征波长（OWs1和OWs2），以及利用灰度共生矩阵（GLCM）提取纹理特征（TFs）。使用植物营养测量仪测量梨树叶片的叶绿素含量（SPAD），并作为梨树叶片黑斑病严重程度的评判依据。将光谱特征、TFs与SPAD以不同组合方式作为输入变量，基于径向基函数神经网络（RBFNN）、卷积神经网络（CNN）模型和基于卷积?长短期记忆神经网络结合注意力机制（CNN?LSTM?Attention, CLATT）模型，分别构建梨树叶片黑斑病病害分级模型，并对检测结果进行分析。测试集分类结果表明，基于OWs1+TFs+SPAD组合作为特征变量输入构建的CLATT模型分类识别精度最高，平均准确率为98.63%，其中健康、轻度、中度和重度梨树叶片样本检测准确率分别可达98.00%、98.21%、99.24%和99.06%。研究结果可为实现梨树叶片黑斑病病害分级检测提供新思路，并指导精准施药。

Abstract:

Pear leaf spot disease, caused by fungi of the genus Alternaria, is a widespread infectious disease that severely impacts the growth and yield of pear trees. Qualitative analysis of the infection severity of pear leaf spot disease, coupled with precise pesticide application based on this assessment, is of paramount importance for reducing economic losses and achieving a reduction in pesticide usage. The hyperspectral imaging technology was employed to analyze different types of pear leaves—-healthy, mild, moderate, and severe leaves. By integrating spectral, image, and chlorophyll features, a high?precision disease grading model was developed to accurately classify and assess the severity of black spot disease on pear leaves. Hyperspectral data were collected by using a specialized imaging system. The spectral data were preprocessed with Savitzky?Golay (SG) smoothing, and characteristic wavelengths were selected through the successive projection algorithm (SPA) and competitive adaptive reweighted sampling (CARS), which identified the optimal wavelengths (OWs1 and OWs2). Texture features (TFs) were extracted by using gray?level co?occurrence matrix (GLCM) analysis. chlorophyll content (SPAD) was measured with a plant nutrition meter and served as an indicator of disease severity. Different combinations of spectral, texture, and chlorophyll features were used as input variables to build disease grading models based on radial basis function neural networks (RBFNN), convolutional neural networks (CNN), and a hybrid CNN?LSTM model incorporating an attention mechanism (CLATT). The models? performances were compared and analyzed. The results from the test set showed that the CLATT model, which utilized the combined features of OWs1, TFs, and SPAD, achieved the highest classification accuracy, with an average recognition rate of 98.63%. The detection accuracies for healthy, mild, moderate, and severe leaf samples reached 98.00%, 98.21%, 99.24%, and 99.06%, respectively. The research results can provide ideas for grading detection of pear leaf black spot disease and guide the implementation of accurate pesticide application.