【摘要】 以水稻为研究对象,获取波长400～1 000 nm范围内的水稻冠层高光谱反射率。采用Savitzky-Golay卷积平滑方法对高光谱数据进行预处理,并通过连续投影算法(Successive projections algorithm,SPA)选择特征波长。在此基础上,提出了一种基于多目标鲸鱼优化算法(Non-dominated Sorting whale optimization algorithm,NSWOA)优化的极限学习机(Extreme learning machine,ELM)模型,用于反演水稻冠层氮素含量。利用误差反向传播神经网络(Back propagation neural network,BPNN)和ELM模型,与NSWOA优化后的ELM模型进行对比。结果表明,SPA算法筛选出的特征波长为400、440、487、542、589、660、675、739、766、808、878、912、949 nm。使用筛选后的特征波长反射率构建NSWOA-ELM水稻冠层氮素含量反演模型效果最好,训练集R~2为0.859 3,RMSE为0.200 2 mg/g;验证集R~2为0.854 3,RMSE为0.206 9 mg/g。与BP神经网络和ELM模型相比,NSWOA-ELM在预测能力和模型稳定性方面具有显著优势。综上,基于NSWOA-ELM的水稻冠层氮素含量反演模型能够为水稻生长状况的描述及精准施肥提供可靠支持。更多还原

【Abstract】 Rice was selected as the research subject,and hyperspectral reflectance data of the rice canopy within the range of 400～1 000 nm was collected experimentally.To preprocess this data,the Savitzky-Golay smoothing method was applied,followed by the successive projections algorithm(SPA)to identify characteristic wavelengths.Based on the processed spectral data,an extreme learning machine(ELM) model optimized by the non-dominated sorting whale optimization algorithm(NSWOA) was developed to estimate the nitrogen content in the rice canopy.To evaluate its performance,the NSWOA optimized ELM model was compared with a traditional back propagation neural network(BPNN) and a standard ELM model.The results indicated that the characteristic wavelengths identified by the SPA algorithm were 400 nm,440 nm,487 nm,542 nm,589 nm,660 nm,675 nm,739 nm,766 nm,808 nm,878 nm,912 nm and 949 nm.The NSWOA-ELM model based on reflectance at these selected wavelengths performed best,achieving a determination coefficient(R2) of 0.859 3 for the training set and 0.854 3 for the validation set,with root mean square errors(RMSE) of 0.200 2 mg/g and0.206 9 mg/g,respectively.Compared with the BPNN and standard ELM models,the NSWOA-ELM model demonstrated superior predictive accuracy and model stability.In conclusion and generalization,the NSWOA-ELM rice canopy nitrogen inversion model provided a reliable approach for assessing rice growth conditions and supporting precision fertilization strategies.更多还原