基于全谱数据挖掘技术的土壤有机质高光谱预测建模研究

Using Different Data Mining Algorithms to Predict Soil Organic Matter Based on Visible-Near Infrared Spectroscopy

可见/近红外高光谱技术与建模方法是当前土壤近地传感器研究领域的重要方向,可应用于土壤养分信息的快速获取和农田作物的精确施肥管理。以浙江省水稻土为研究对象,利用以非线性模型为核心的数据挖掘技术,包括随机森林、支持向量机、人工神经网络等方法分别建立了不同建模集和验证集的原始光谱与有机质含量的估测模型。结果表明:研究比较的1∶1,3∶1和全部样本建模并全部验证的三种样本模式划分对建模的结果有一定的影响。相较于目前常用的偏最小二乘回归(PLSR)建模方法而言,非线性模型RF和SVM也取得了较好的建模精度,三种模式下其RDP值均大于1.4。特别是采用SVM建模方法所得模型具有很好的预测能力,模式二下其RDP值达到2.16。同时引入ANN方法改进建立的PLSR-ANN方法显著提高了PLSR的模型预测能力。

Using visible/near infrared spectroscopy to model soil properties is very important in current soil sensing research.It can be applied to rapidly access soil information and precision management.In the present study,paddy soil in Zhejiang Province is treated as the research samples.The nonlinear models such as random forests（RF）,supported vector machines（SVM） and artificial neural networks（ANN） were used respectively to build models to predict soil organic matter based on different selection of calibration and validation datasets.The results show that there is a certain impact on prediction results under the division of different sample modes.Compared to the commonly used linear model PLSR,the nonlinear model RF and SVM have comparable prediction accuracy,especially predictions by SVM using all Vis-NIR wavelengths produced the smallest RMSE values.It shows that the model constructed by SVM method has a good predictive ability.In addition,a combined method,PLSR-ANN（with the introduction of ANN into PLSR）,significantly improves the predictive ability of PLSR.Even though ANNs are ＂black box＂ systems the combination of PLSR and nonliner modelling helps achieve good predictions and interpretability.