Estimating protein content of rice grain based on hyperspectral reflectance data is a useful measure for non-destructively,rapid and accurately monitoring rice grain quality.In this study,field experiments were carried out in Haining City,Zhejiang Province.Rice grain reflectance data and protein contents of rice grains were experimented in harvest time and analysed by single wave correlation,regression model and neural network model.Results:The grain protein content is highly correlative with the spectrum reflectance and differential spectrum reflectance.At blue,red and infrared wavelength the correlation between grain protein content and spectrum reflectance is significant,while the correlation between grain protein content and differential spectrum reflectance is of variety.There are also significant correlations between five vegetation indexes and grain protein content.Correlation coefficients between model prediction values and experimental values range from 0.44 to 0.55,and RMSE range from 0.46 to 0.79.However,due to rather complicated non-linear relations existing between hyperspectral reflectance data and grain protein content,the results of grain protein content retrieved from the statistic model are not so ideal.For this reason,an artificial neural network model (BP model) was constructed and applied in the retrieval of grain protein content.For its superior ability for solving the non-linear problem,the BP model provides a much better accuracy in retrieval of grain protein content compared with the results from the statistic model.It is implied that grain protein content could be predicted by remote sensing technology.%探索不同高光谱模型估测水稻籽粒蛋白质含量的精度,对于快速、无损、准确地监测水稻籽粒品质具有重要意义.在浙江省海宁市晚稻收获期同步采集了水稻籽粒光谱反射率与蛋白质含量数据,并分别对获取的数据进行了单波段相关分析、植被指数回归分析、神经网络模拟.通过分析得出以下结论:(1)水稻籽粒光谱反射率、微分光谱与对应的蛋白质含量密切相关.其中,在蓝光、红光和近红外处光谱反射率与水稻籽粒蛋白质含量的相关系数和确定性系数稳定且较高,而微分光谱与水稻籽粒蛋白质含量之间的相关系数和确定性系数变化较大,在500～750 nm间则相对稳定;(2)5种植被指数和水稻籽粒蛋白质含量均密切相关.线性模型的预测值与实测值之间的相关系数在0.44～0.55之间,均方根差在0.46～0.79之间;(3)神经网络模型可以容纳更多的相关波段参与水稻籽粒蛋白质含量的估算,实测值与预测值的相关系数高于其他模型,而均方根差低于其他模型,可以用于快速无损检测水稻籽粒蛋白质含量.
展开▼
