基于多源特征融合的玉米大螟危害等级监测研究

Monitoring of Asian Corn Borer Damage Levels Based on Multi-source Feature Fusion

摘要:

大螟在玉米生长早期造成的茎秆破坏切断了水分和养分运输，该虫害无损精准检测技术的应用对优化防控策略、提升玉米生产效益具有显著性影响。本研究提出一种基于多源特征融合的玉米大螟危害等级（Asian corn borer damage levels，ACBDL）监测方法，融合玉米三叶期植被指数、纹理特征和颜色指数，构建玉米早期大螟危害等级监测模型，提高玉米大螟危害等级预测精度。利用无人机搭载的RGB及多光谱成像系统采集玉米三叶期光谱影像，采用监督分类中的马氏距离分类（Mahalanobis distance classification，MDC）将玉米与土壤进行分类，并进行二值化掩膜剔除土壤背景。提取过量绿色指数（Excess green index，ExG）和土壤调节植被指数（Soil-adjusted vegetation index，SAVI）等14种植被指数，基于灰度共生矩阵（Gray-level co-occurrence matrix，GLCM）计算4个波段共32种纹理特征，转化计算得到8种颜色特征参数。采用皮尔逊相关系数法（Pearson correlation coefficient, PCC）筛选特征，构建机器学习模型随机森林（Random forest，RF）、极端梯度提升（Extreme gradient boosting，XGBoost）、K最近邻（K-nearest neighbors，KNN）和类别提升（Categorical boosting，CatBoost）预测模型。结果显示：多源特征融合可显著提高模型预测精度，KNN模型在植被指数、纹理特征和颜色指数融合的条件下综合性能表现最优，总体准确率、精确率、召回率、F1值和Kappa系数分别为91.8%、91.9%、91.8%、89.5%和87.4%。该研究验证了多源特征融合在大螟危害等级预测中的有效性，为玉米早期病虫害防治提供可靠技术参考。

Abstract:

The Asian corn borer causes stem damage in the early growth stage of maize, disrupting the transport of water and nutrients. The application of non-destructive and precise detection techniques is crucial for optimizing pest control strategies and improving maize production efficiency. A multi-source feature fusion-based method for monitoring the corn borer damage level (CBDL) was proposed, integrating vegetation indices, texture features, and color indices of maize at the three-leaf stage to enhance the overall accuracy of early-stage damage assessment. UAV-mounted RGB and multispectral imaging systems were employed to acquire spectral data during the three-leaf stage. The mahalanobis distance classification (MDC) algorithm under supervised classification was used to distinguish maize from soil, followed by binary masking to remove soil background. Fourteen vegetation indices, including the excess green index (ExG) and soil-adjusted vegetation index (SAVI) were extracted;totally 32 texture features were computed from four bands based on the gray-level co-occurrence matrix (GLCM);and eight color parameters were derived. Features were selected by using the Pearson correlation coefficient (PCC), and machine learning prediction models, including random forest (RF), extreme gradient boosting (XGBoost), K-nearest neighbors (KNN), and categorical boosting (CatBoost) were constructed. Results indicated that multi-source feature fusion significantly improved model prediction overall accuracy. Among all models, the KNN model integrating vegetation, texture, and color features achieved the best overall performance, with an overall accuracy, precision, recall, F1-score, and Kappa coefficient of 91.8%, 91.9%, 91.8%, 89.5%, and 87.4%, respectively. The findings demonstrated the effectiveness of multi-source feature fusion in predicting the damage level of corn borer infestations, and it can provide a reliable technical reference for early detection and control of maize pests.