基于多生育期光谱变量的水稻直链淀粉含量监测

直链淀粉含量是评价稻米品质的重要指标之一,其累积生长过程是多生育期、多因素综合作用的结果。为了探究多生育期信息引入对水稻籽粒直链淀粉含量监测模型的影响,实现水稻品质信息的大规模准确监测。该研究选取水稻孕穗期、抽穗期、灌浆期和成熟期这4个有关水稻籽粒形成发育的生育期的冠层光谱,分析原光谱、植被指数、高光谱特征参数,及其变换形式与水稻籽粒直链淀粉含量的相关性,筛选得到相关性较好的光谱变量,并利用逐步回归的方法进行建模,建立基于多生育期光谱变量的直链淀粉含量预测模型。结果表明:一阶导数、差值植被指数(Difference Vegetation Index,DVI)、比值植被指数(Ratio Vegetation Index, RVI)及成熟期特征参数表现出较高敏感性,最适用于直链淀粉含量预测的生育期为成熟期,而多生育期信息的综合利用能显著提高模型预测精度,最佳多生育期预测模型为孕穗-抽穗-成熟期组合模型,建模决定系数(Coefficient of Determination, R^2)为0.708,均方根误差(Root Mean Square Error, RMSE)为0.711%,平均绝对百分比误差(Mean Absolute Percent Error, MAPE)为3.22%,验证R2为0.631,RMSE为0.768%,MAPE为3.99%,证明该模型能较为精确地预测籽粒直链淀粉含量,为稻米品质指标大尺度统计监测提供一定的技术支撑和应用基础。

Monitoring of amylose content in rice based on spectral variables at the multiple growth stages

Amylose content is one of the important indexes for evaluating rice quality. Large-scale and rapid monitoring of rice quality is of great significance for measuring rice commodity value and regulating food crop production. Because amylose is wrapped in rice grains and cannot be directly expressed in the canopy spectrum, the accuracy of its canopy prediction model is often low. Considering that the accumulation and growth of rice starch granules are the result of multiple growth stages and multiple factors, this study attempts to introduce multiple growth stages information to improve the model, while most existing studies only use single growth stage information. The research area was located in Deqing County, Zhejiang Province, China The experiment spanned two rice growing seasons from 2016 to 2017, with five nitrogen levels and three rice varieties. Correlation relationships between the original spectra and first derivative spectra of rice canopy at booting stage, heading stage, milking stage and maturity stage and the grain amylose content were analyzed, then four types of vegetation indices and 23 hyperspectral features for further correlation analysis were computed. According to results of correlation analysis, the suitable spectral variables with high correlation coefficient were selected for amylose content modeling by stepwise regression method. The prediction models were established for different single growth stages to obtain the best growth stage of amylose prediction. Then, by combining the information of different growth stages, the amylose content prediction models based on the combination of different growth stages were established, and the effect of comprehensive application of multiple growth stage information on the amylose content prediction model was analyzed to get the best prediction model and its growth stages combination. The results showed that the first derivative, Difference Vegetation Index(DVI), Ratio Vegetation Index(RVI) and the hyperspectral features at maturity stage were highly sensitive to amylose content. The derivative of 1 649 nm and 1 610 nm showed a good explanation for amylose content, 1 600-1 700 nm might be the sensitive sepctral bands of rice amylose prediction. In addition, the characteristic parameters of maturity-stage spectrum showed a strong explanatory ability in the maturity-stage model of this study, especially the blue edge position(λb), but this variable rarely appeared in other related research prediction models, and its principle and stability with strong prediction ability in maturity-stage model need further study and verification. The results of single growth stage modeling showed that the accuracy of the maturity and heading stages models was significantly higher than that of booting and milking stages,the most suitable growth stage for predicting amylose content was maturity stage, with the modeling coefficient of determination(R^2)=0.558, Root Mean Square Error(RMSE)=0.896%, Mean Absolute Percent Error(MAPE)=4.49%, and validation R^2=0.629, RMSE=0.864%, MAPE=4.59%. The comprehensive utilization of multi-growth stage information could significantly improve the prediction accuracy of the model, and the best multi-growth stage prediction model was the combination model of booting-heading-maturity stage, with the modeling R^2=0.708, RMSE=0.711%, MAPE=3.22%, and validation R^2=0.631, RMSE=0.768%, MAPE=3.99%, which proved that the model could accurately predict amylose content in grains.