基于无人机成像高光谱影像的冬小麦LAI估测

利用无人机Cubert UHD185 Firefly成像光谱仪和ASD光谱仪获取了冬小麦挑旗期、开花期和灌浆期的成像和非成像高光谱以及LAI数据。首先,对比ASD与UHD185光谱仪数据光谱反射率,评价两者精度;然后,选取7个光谱参数,分析其与冬小麦3个生育期LAI的相关性,并使用线性回归和指数回归挑选出最佳估测参数;最后利用多元线性回归、偏最小二乘、随机森林、人工神经网络和支持向量机构建了冬小麦3个不同生育期LAI的估测模型。结果表明:UHD185光谱仪光谱反射率在红边区域与ASD光谱仪趋势一致性很高,反射率在挑旗期、开花期、灌浆期的R^2分别为0.9959、0.9990和0.9968,UHD185光谱仪数据精度较高;7种光谱参数在挑旗期、开花期、灌浆期与LAI相关性最高的参数分别是NDVI(r=0.738)、SR(r=0.819)、NDVI×SR(r=0.835);LAI-MLR为冬小麦LAI的最佳估测模型,其中开花期拟合性最好,精度最高(建模R^2=0.6788、RMSE为0.69、NRMSE为19.79%,验证R^2=0.8462、RMSE为0.47、NRMSE为16.04%)。

Leaf Area Index Estimation of Winter Wheat Based on UAV Imaging Hyperspectral Imagery

The UHD185 imaging spectrometer and ASD spectroradiometer were used to acquire imaging and non imaging hyperspectral data during three wheat growth stages, including flagging stage, flowering stage and filling stage. The corresponding ground leaf area index (LAI) data were also collected. Firstly, the ASD and the UHD185 spectrometer data were compared and their precision was evaluated. Then, the correlation analyses were conducted between LAI and seven LAI related spectral parameters, linear regression and exponential regression were used to select the optimal estimation parameters. Finally, for each growth stage, multivariate linear regression, partial least squares, random forest, artificial neural network and support vector machine were used to construct LAI estimation models for winter wheat. The experimental results showed that UHD185 hyperspectral spectrometer reflectance was highly consistent with ASD ground hyperspectral spectrometer reflectance in the red edge region. The coefficients of determination between them were 0.9959, 0.9990 and 0.9968 for flagging stage, flowering stage and filling stages, respectively. The parameters with the highest correlation with LAI were NDVI (r=0.738) for flagging stage, SR (r=0.819) for flowering stage, and NDVI×SR (r=0.835) for filling stage. LAI-MLR was the best estimation model for winter wheat. The highest accuracy for flowering stage with R2 of 0.6788, RMSE of 0.69 and NRMSE of 19.79% for calibration, and with R2 of 0.8462, RMSE of 0.47 and NRMSE of 16.04% for validation.