基于特征波段选择的冬小麦叶面积指数高光谱遥感估测模型研究

Research on Hyperspectral Remote Sensing Estimation Model of Winter Wheat Leaf Area Index Based on Feature Band Selection

为提高冬小麦叶面积指数(LAI)的遥感估测精度,以实现其无损快速测定目标,在田块尺度设置多年定点不同冬小麦品种氮梯度试验,测定其不同生育时期冠层高光谱数据和LAI,通过原始冠层光谱数据与一阶导数预处理(first-derivative, FD)组合竞争自适应重加权采样(competitive adaptive reweighted sampling, CARS)、无信息变量消除(uninformative variable elimination, UVE)和随机蛙跳(random frog, RF)三种特征波段选择方法进行偏最小二乘回归(partial least squares regression, PLSR)高光谱估测模型构建。结果表明,一阶导数预处理在简化波段数量和提升模型精度上具有较好作用。经过与全波段数据及六种组合内部建模预测精度对比,RF在简化波段方面效果最好,FD-RF组合筛选波段数量为6个,建模的R^(2)和RMSE分别达到0.850和0.730,预测的R^(2)和RMSE分别为0.704和1.005;FD-CARS组合达到了最佳建模精度,R^(2)和RMSE分别为0.876和0.641;FD-UVE组合达到了最佳预测精度,R^(2)和RMSE分别为0.755和0.672。这说明基于特征波段选择可以进行冬小麦叶面积指数高光谱遥感模型建立与有效估测。

In order to improve the remote sensing estimation accuracy of the leaf area index(LAI)of winter wheat and achieve its non-destructive and rapid measurement goal,nitrogen gradient experiments were conducted on different winter wheat varieties at field scale for many years,and canopy hyperspectral data and LAI were measured at different growth stages.The partial least squares regression(PLSR)hyperspectral estimation model was constructed by competing with the original canopy spectral data and first-derivative(FD)preprocessing methods of competitive adaptive reweighted sampling(CARS),uninformative variable elimination(UVE)and random frog(RF).The results showed that the first-derivative preprocessing had an optimal performance in simplifying the number of bands and improving the accuracy of the model.Compared with the prediction accuracy of the full-band data and the six combinations,RF had the strongest performance in terms of simplified bands among which the number of FD-RF combination screening bands was six;the modeled R^(2) and RMSE reached 0.850 and 0.730,respectively,and the predicted R^(2) and RMSE were 0.704 and 1.005,respectively.Moreover,the FD-CARS combination achieved the best modeling accuracy,with R^(2) and RMSE reaching 0.876 and 0.641,respectively.The FD-UVE combination achieved the best prediction accuracy,with R^(2) and RMSE reaching 0.755 and 0.672,respectively.It suggested that the hyperspectral remote sensing model of winter wheat leaf area index could be established and estimated effectively based on the selection of characteristic bands.