基于无人机RGB影像的小麦叶面积指数与产量估算研究

Estimation of Wheat Leaf Area Index and Yield Based on UAV RGB Images

为探讨基于无人机RGB影像实现对小麦叶面积指数(leaf area index,LAI)和产量估算的可行性,设置不同生态点、品种和氮素处理的小麦田间试验,应用大疆精灵4 Pro无人机获取小麦拔节期、抽穗期、扬花期和灌浆期4个主要生育时期的RGB高时空分辨率影像,并同测定小麦LAI。采用相关性分析筛选出不同生育时期对LAI敏感的光谱与纹理特征集,并借助随机森林(random forest,RF)、偏最小二乘回归法(partial least squares regression,PLSR)、BP神经网络(back propagation neural network,BPNN)和支持向量机(support vector machine,SVM)分析方法,筛选出小麦不同生育时期最优的LAI估测模型。基于不同生育时期的光谱与纹理特征以及时期特征集,进一步建立产量预测模型,并在不同生态点验证叶面积估算模型与产量预测模型的普适性。结果表明,基于RF的LAI估测模型的验证精度最高,4个生育时期的均方根误差(root mean square error,RMSE)分别为2.26、1.44、1.73和1.02。基于RF的产量预测模型验证效果也最优,RMSE为1.17 t·hm^(-2)。由此说明基于无人机RGB影像和RF算法,建立LAI和产量估测模型,可实现小麦长势实时监测和产量预测。

In order to explore the feasibility of estimating leaf area index(LAI)and yield of wheat based on RGB images of UAVs,field experiments were conducted with different wheat varieties and nitrogen treatments at different ecological sites.DJI Phantom 4 Pro UAV was employed to obtain RGB images with high spatial-temporal resolution at the stage of jointing,heading,anthesis and grain-filling stages.LAI was synchronously measured.The spectral and texture feature sets sensitive to at different growth stages were screened by correlation analysis,and the random forest(RF),partial least squares regression(PLSR),back propagation neural network(BPNN)and support vector machine(SVM)analysis methods were used.The optimal LAI estimation models at different growth stages of wheat were selected.Further more,the optimal yield prediction was modeled based on the features of spectral,texture and LAI at different growth stages.Both the LAI estimation model and yield prediction model were validated at different locations.The results showed that the RF-based LAI estimation model had the highest validation accuracy,with the root mean square error(RMSE)at different growth stages of 2.26,1.44,1.73,and 1.02,respectively.Also,the RF-based yield prediction model had the best prediction accuracy,with the RMSE of 1.17 t·hm^(-2).Based on the UAV platform,the wheat LAI monitoring model and yield prediction model constructed in this study will provide effective technology for the real-time monitoring of wheat growth and yield prediction.