摘要
【目的】条锈病对小麦生长和产量造成严重威胁,为确保有效防控,精准监测尤为关键。利用遥感技术构建小麦条锈病估测模型,能快速、准确地估测病情指数(DI),为精准防控提供技术支持。【方法】利用ASD光谱仪获取小麦不同生育期(抽穗期、灌浆期和成熟期)高光谱数据,采用随机森林变量选择(VSURF)方法结合相关性分析(CA)对原始光谱(OR)和一阶微分光谱(FD)进行特征波段筛选。使用随机森林(RF)对比不同数据集的特征波段建模结果,确定模型效果最佳的特征集。随后借助偏最小二乘回归(PLSR)、极致梯度提升(XGBoost)以及反向传播神经网络(BPNN),对比特征集在不同算法中的建模效果。通过对比建模效果,确定针对全生育期小麦条锈病病情指数的最佳估测模型。为了验证特征集在不同生育期中的效果,利用特征集在3个生育期重新构建模型,并对比模型效果。【结果】对不同数据集进行特征筛选,并使用RF构建条锈病DI估测模型,通过比较模型效果,确定VSURF-CA-FD特征集(绿光范围的537 nm以及近红外范围的821和846nm)在RF模型中的估测效果最好。采用RF算法构建的模型表现出优异的精度,R^(2)为0.89,RMSE为12.34。这些特征波段在其他算法构建的模型中也展现出良好的精度:XGBoost模型的R^(2)为0.87,RMSE为13.15;BPNN模型的R^(2)为0.84,RMSE为15.19;PLSR模型的R^(2)为0.69,RMSE为20.92。使用不同生育期的冠层微分高光谱数据进行验证,利用VSURF-CA-FD特征集构建RF模型,对比模型发现在小麦生长的早期(抽穗期)R^(2)为0.54,RMSE为1.29,NRMSE为0.21,能满足估测病害的要求;小麦生长的中期(灌浆期),模型的R^(2)表现较好,R^(2)为0.66,RMSE为12.24,NRMSE为0.21;小麦生长晚期(成熟期),模型效果好于前两个时期,R^(2)为0.75,RMSE为10.77,NRMSE为0.15。【结论】使用VSURF-CA方法筛选出的特征波段,能构建出对小麦条锈病病情指数具有出色估测效果的RF模型。研究结果可为预测早期和中期条锈病病情指数提供有价值的思路和方法。
【Objective】Stripe rust is a serious threat to the growth and yield of wheat.Accurate monitoring and diagnostic assessment are fundamental prerequisites for effective prevention and control of stripe rust.The objective of this study is to construct a wheat stripe rust estimation model using remote sensing technology,enable the rapid and precise estimation of the disease index(DI),and to provide technical support for precise prevention and control.【Method】The hyperspectral data of wheat at different growth stages(heading period,grain-filling period,and maturity period)were acquired through the ASD spectrometer.Initially,the variable selection using random forests(VSURF)method,combined with correlation analysis(CA),was applied to select characteristic bands from the original spectrum(OR)and the first-order differential spectrum(FD).Subsequently,the random forest(RF)algorithm was utilized to compare modeling results of characteristic bands from different datasets,identifying the feature set with the most effective model.Further,models such as partial least squares regression(PLSR),extreme gradient boosting(XGBoost),and back-propagation neural network(BPNN)were employed to compare the modeling effects of different feature sets within various algorithms.This comprehensive analysis aimed to determine the optimal estimation model for wheat stripe rust DI across the entire growth period.Simultaneously,to validate the effectiveness of the feature set across different growth stages,the feature set was used to rebuild models during each of the three distinct growth periods.【Result】The comparative analysis of model effects revealed that the VSURF-CA-FD feature set(537 nm in the green range and 821,846 nm in the near-infrared range)demonstrated the most effective estimation within the RF model,achieving an R^(2)value of 0.89 and an RMSE of 12.34.These feature bands also exhibited precision in models constructed with other algorithms,including XGBoost(R^(2):0.87,RMSE:13.15),BPNN(R^(2):0.84,RMSE:15.19),and PLSR(R^(2):0.69,RMSE:20.92).For models constructed during different growth stages,the early growth stage(heading period)exhibited an R^(2)value of 0.54,RMSE of 1.29,and NRMSE of 0.21,meeting the requirements for disease estimation.In the middle growth stage(grain-filling period),the model performed well with an R^(2)of 0.66,RMSE of 12.24,and NRMSE of 0.21.In the late growth stage(maturity period),the model’s effectiveness surpassed that of the previous two stages,with an R^(2)of 0.75,RMSE of 10.77,and NRMSE of 0.15.【Conclusion】Utilizing characteristic bands selected through the VSURF-CA method,an RF model with excellent estimation accuracy for wheat stripe rust DI can be established.The research outcomes will provide valuable insights and methodologies for predicting early and mid-stage stripe rust DI.
作者
梅广源
李荣
梅新
陈日强
樊意广
程金鹏
冯子恒
陶婷
赵倩
赵培钦
杨小冬
MEI GuangYuan;LI Rong;MEI Xin;CHEN RiQiang;FAN YiGuang;CHENG JinPeng;FENG ZiHeng;TAO Ting;ZHAO Qian;ZHAO PeiQin;YANG XiaoDong(Information Technology Research Center,Beijing Academy of Agriculture and Forestry Sciences/Key Laboratory of Quantitative Remote Sensing in Agriculture,Ministry of Agriculture and Rural Affairs,Beijing 100097;Faculty of Resources and Environmental Science,Hubei University,Wuhan 430062)
出处
《中国农业科学》
CAS
CSCD
北大核心
2024年第3期484-499,共16页
Scientia Agricultura Sinica
基金
国家重点研发计划(2023YFD2000105)
国家自然科学基金(41771469)。
