基于分数阶微分和最优光谱指数的大豆叶面积指数估算

高光谱遥感技术可对作物生长状况进行无损、高效地监测，是推动现代精准农业发展的必要手段。以不同施氮水平与覆膜处理下的开花期大豆叶面积指数(Leaf area index, LAI)为研究对象，对原始开花期大豆高光谱反射率数据进行0～2阶微分变换处理(步长0.5),并筛选出各阶光谱指数中与开花期大豆LAI相关性最高的指数作为最优光谱指数进行输入，采用支持向量机(Support vector machine, SVM)、随机森林(Random forest, RF)、遗传算法优化的BP神经网络(BP neural network optimized by genetic algorithm, GA-BP)3种机器学习方法构建大豆LAI预测模型。结果表明：0～2阶光谱指数与大豆LAI相关系数平均值分别为0.616、0.657、0.666、0.669、0.658,相比于原始与整数阶高光谱反射率，分数阶微分变换处理后的高光谱反射率构建的光谱指数与开花期大豆LAI具有更强的相关性；相关系数平均值最高的1.5阶微分处理最优光谱指数波长组合分别为：TVI(687 nm, 754 nm)、DI(687 n...

Estimation of Leaf Area Index of Soybean Based on Fractional Order Differentiation and Optimal Spectral Index

Hyperspectral remote sensing crop growth monitoring technology, an essential instrument for developing contemporary precision agriculture, is characterized by non-destructiveness and real-time effectiveness. Taking leaf area index (LAI) of soybean at flowering stage under different levels of N application and mulching treatment as research object, the raw data for the hyperspectral reflectance of the soybean canopy were pretreated by using the 0~2 order differential transform processing (step 0.5). Based on five sets of pretreatment reflectance data, the optimum spectral index with a high correlation to the LAI of soybean at the blooming stage was the input data. And the support vector machine (SVM), random forest (RF), and BP neural network optimized by genetic algorithm (GA-BP) were used to construct the soybean LAI prediction model.The results showed that compared with the integer order and the raw hyperspectral reflectance, the spectral indices built from the fractional order differential preprocessed hyperspectral reflectance correlated better with the soybean LAI.The corresponding bands of different orders of optimal spectral indices concentrated in the red-edge band. The correlation between the spectral index and soybean LAI was increased and then decreased as the differential order was increased, and the accuracy of the prediction model showed the same pattern.When the input data were the same for all three machine learning techniques, the model created by RF had the highest accuracy. A thorough analysis determined that the soybean LAI prediction model built by using RF had the highest accuracy of prediction when the input variable was the 1.5-order differential optimal spectral index. The R2 of the model validation set was 0.880, the RMSE was 0.3200cm2/cm2, the NRMSE was 10.354% and the MRE was 9.572%. The research result can help advance the development of precision agricultural production by offering theoretical references for enhancing the inversion accuracy of soybean LAI hyperspectral prediction models.