基于互信息理论的水稻磷素营养高光谱诊断

当前，磷素营养诊断的化学分析方法既费力又费时，使诊断结果难以及时应用到田间生产，而高光谱遥感技术是一种非破坏性、快速和有潜力的作物营养诊断技术。但是，由于光谱分析技术的限制，作物磷营养与光谱特性之间的关系研究进展一直较为缓慢。文章通过室内实验获取了不同磷营养水平水稻典型生育期冠层光谱反射率及其对应的磷、叶绿素含量等农学参数，并对农学参数做了LSD多重比较。利用互信息(MI)理论分析了水稻磷素含量的敏感波段，结果表明水稻拔节期叶片磷素估测的敏感波段分别为536，630，1 040，551和656 nm，与其相对应的互信息值分别为1.057 5，1.103 9，1.135 3，1.141 7和1.149 4;比较了以此敏感波段为自变量构建的BP人工神经网络模型和多元线性回归模型，结果显示BP人工神经网络模型更优，其交叉验证均方根误差(RMSE-train)和相关系数(R2)分别为0.038 8和0.988 2，而预测均方根误差(RMSE-test)和相关系数(R2)分别为0.050 5和0.989 2。说明利用互信息-神经网络模型(MI-ANN)和高光谱遥感估测田间水稻磷含量是可能的。

Application of Mutual Information to Variable Selection in Diagnosis of Phosphorus Nutrition in Rice

The present study obtained data of rice canopy spectrum, and P and chlorophyll content at typical growth stages with different rates of P supply by means of solution experiment. The effects of P treatments on leaf P and chlorophyll content were analyzed statistically using LSD’s multiple comparison at a probability of 0.05; By mutual information (MI) variable selection procedure, the optimal spectral variables were identified at 536, 630, 1 040, 551 and 656 nm, and their corresponding mutual information values were 1.057 5, 1.103 9, 1.135 3, 1.141 7 and 1.149 4 respectively; based on these sensitive bands, the built feed-forward artificial neural network model (ANN) had higher precision for P content estimation than the multiple linear regression model (MLR). Its RMSE of cross-validation and R were 0.038 8 and 0.988 2, respectively, for the calibration data set, and the RMSE of prediction and R were 0.050 5 and 0.989 2, respectively, for the test data set. Therefore, it was suggested that MI was encouraged for quantitative prediction of leaf P content in rice with visible/near infrared hyperspectral information without assumption on the relationship between independent and dependent variables. But more work is needed to explain why these bands are sensitive to leaf P content in rice.