Real-time or pre-symptomatic wheat scab (WS) detection is inevitable for precision agriculture to secure yield and quality at the critical grain formation stage. For this, feature selection (FS) techniques and machine learning (ML) have demonstrated their capabilities. However, for the same type and size of dataset, all FS and ML techniques behave differently due to their diverse primary constituents. This study attempts to leverage ML for WS classification and prediction employing different FS techniques on hyperspectral data of wheat spikes. The spectral features were selected and assessed to regress and classify disease occurrence. Relief-F-neural net (NN) manifested the best results with classification accuracy (CA) of 67 % and 89 % at the pre-symptomatic scale and 3 days after inoculation (DAI), respectively. Followed by continuous wavelet transform (CWT)-NN with 63 % CA at the pre-symptomatic scale and CWT-Xgboost with 89 % CA at 3DAI. For prediction, random forest regression revealed best accuracy of R2 = 0.94 and RMSE = 7.70, followed by partial least squares regression with R2 = 0.90 and RMSE = 10.37. The results offer a precise quantitative benchmark for future investigations into the capacity of hyperspectral data and FS for the real-time quantification of plant diseases.

中文翻译：

---

利用机器学习，通过高光谱反射和特征选择方法区分小麦赤霉病感染水平


实时或症状前小麦赤霉病 （WS） 检测对于精准农业来说是不可避免的，以确保关键谷物形成阶段的产量和质量。为此，特征选择 （FS） 技术和机器学习 （ML） 已经展示了它们的功能。但是，对于相同类型和大小的数据集，由于 FS 和 ML 技术的主要成分不同，其行为也不同。本研究试图利用 ML 对小麦穗的高光谱数据采用不同的 FS 技术进行 WS 分类和预测。选择并评估光谱特征以回归和分类疾病发生。Relief-F-neural net （NN） 在症状前量表和接种后 3 天 （DAI） 的分类准确率 （CA） 分别为 67% 和 89% 的最佳结果。其次是症状前尺度上 CA 为 63% 的连续小波变换 （CWT）-NN 和 3DAI 时 89% CA 的 CWT-Xgboost。对于预测，随机森林回归显示 R2 = 0.94 和 RMSE = 7.70 的最佳准确性，其次是偏最小二乘回归，R2 = 0.90 和 RMSE = 10.37。这些结果为未来研究高光谱数据和 FS 实时量化植物病害的能力提供了精确的定量基准。