应用局部神经网络和可见/近红外光谱法估测土壤有效氮磷钾

要实现农田合理施肥，需要对土壤养分状况进行实时、准确地诊断，因而建立快速、稳定可靠的土壤养分定量分析方法是关键。光谱分析是一种有很大潜力的快速分析方法，从可见/近红外光谱建模的几个重要环节，即特征波段、预处理方法及回归模型方法的选择，研究了土壤有效氮、磷、钾含量快速估测的光谱建模方法。采用了多元散射校正加一阶导数进行光谱预处理，通过逐波段相关分析在可见-近红外区优选特征波段，并应用了局部非线性回归方法(BP神经网络局部回归法)建模，所建模型对土壤有效氮、磷、钾含量估测的相关系数r分别为0.90，0.82和0.94，BP神经网络局部建模比全局建模具有更好的精度和稳定性，估测精度提高幅度分别为40.63%，28.64%，22.90%。因此，采用局部BP神经网络回归建模法建立土壤有效氮、磷、钾的光谱定量分析模型，可实现对土壤养分状况的快速诊断。该研究的创新点是通过采用局部非线性回归方法提高了土壤光谱营养诊断模型的稳定性和可靠性，为作物生长过程中不同生长时期的土壤养分的动态监测和过程控制提供了技术支持。

Applying Local Neural Network and Visible/Near-Infrared Spectroscopy to Estimating Available Nitrogen,Phosphorus and Potassium in Soil

To establish the quantitative relationship between soil spectrum and the concentration of available nitrogen, phosphorus and potassium in soil, the critical procedures of a new analysis method were examined, involving spectral preprocessing, wavebands selection and adoption of regression methods. As a result, a soil spectral analysis model was built using VIS/NIRS bands, with multiplicative scatter correction and first-derivative for spectral preprocessing, and local nonlinear regression method (Local regression method of BP neural network). The coefficients of correlation between the chemically determined and the modeled available nitrogen, phosphorus and potassium for predicted samples were 0.90, 0.82 and 0.94, respectively. It is proved that the prediction of local regression method of BP neural network has better accuracy and stability than that of global regression methods. In addition, the estimation accuracy of soil available nitrogen, phosphorus and potassium was increased by 40.63%, 28.64% and 28.64%, respectively. Thus, the quantitative analysis model established by the local regression method of BP neural network could be used to estimate the concentration of available nitrogen, phosphorus and potassium rapidly. It is innovative for using local nonlinear method to improve the stability and reliability of the soil spectrum model for nutrient diagnosis, which provides technical support for dynamic monitoring and process control for the soil nutrient under different growth stages of field-growing crops.