基于光谱特征参数的粳稻冠层氮素含量反演方法

A Study on Inversion Method of Nitrogen Content in Japonica Rice Based on Spectral Characteristic Parameters

粳稻氮素含量的快速、无损、准确估算,可以及时掌握粳稻的生长状况,对指导粳稻田间管理具有重要意义。为提高粳稻冠层氮素含量的高光谱反演精度,利用沈阳农业大学路南试验基地2018年粳稻3个关键生育期无人机高光谱影像和同步测定的粳稻冠层氮素含量作为数据源,选用从粳稻冠层光谱中提取的高光谱位置变量、面积变量和植被指数变量3种类型20个光谱特征参数与氮素含量进行相关性分析,选出各个生育期内相关性较高的前3个光谱特征参数作为模型输入分别建立偏最小二乘回归(PLSR)、BP神经网络(BPNN)和思维进化算法优化BP神经网络(MEA-BPNN)3种粳稻冠层氮素含量反演模型并验证。结果表明:在粳稻分蘖期、拔节期、抽穗期,与粳稻氮素含量相关性最好的高光谱特征参数均为红边面积SDr,相关系数分别为0.771,0.664,0.775;MEA-BPNN反演模型与PLSR、BPNN相比,无论在模型精度还是预测能力都有明显提高,在各个生育期,MEA-BPNN模型的建模集和验证集决定系数R^2均达到0.700以上,RMSE均低于0.400以下,说明MEA-BPNN反演模型是筛选出的最佳粳稻冠层氮素含量反演模型。综上研究,该模型能够快速无损反演粳稻冠层氮素含量,可为后续施肥决策提供支持。

The rapid,non-destructive and accurate estimation of the nitrogen content of Japonica rice can timely grasp the growth status and provide the great significance for guiding the management of Japonica rice field.In order to improve the accuracy of hyperspectral inversion of nitrogen content in japonica rice canopy,the unmanned aerial vehicle(UAV)hyperspectral imagery and simultaneous determination of nitrogen content in canopy of Japonica rice were used as data sources for three key growth stages of Japonica rice in the experimental base of Shenyang Agricultural University in 2018.The hyperspectral position extracted from the canopy spectrum of the Japonica rice was selected.Correlation analysis between the three spectral types of the three types of variables,the area variable and the vegetation index variable and the nitrogen content,and the first three spectral characteristic parameters with high correlation during each growth period were selected as the model inputs.Partial least squares regression(PLSR),BP neural network(BPNN)and thought Optimizing BP Neural Network with Mind Evolutionary Algorithms(MEA-BPNN)and other three models have canopy nitrogen contented inversion and verification with Japonica rice.The results show thatthe best hyperspectral parameters in the tillering,jointing and heading stages of japonica rice were red edge areaSDr,and the correlation coefficients were 0.771,0.664 and 0.775 respectively.Compared with PLSR and BPNN,MEA-BPNN inversion model has obvious improvement in model accuracy and prediction ability.In each growth period,the MEA-BPNN model's modeling set and verification set determination coefficient R2 have reached more than 0.700 and RMSE is below 0.400,which shows that the MEA-BPNN inversion model is the best inversion model of canopy nitrogen content in japonica rice.In summary,the model can quickly and non-destructively invert the nitrogen content of Japonica rice canopy,which can support the subsequent fertilization decision.