基于梯度提升树的土壤速效磷高光谱回归预测方法

Predicting Soil Available Phosphorus by Hyperspectral Regression Method Based on Gradient Boosting Decision Tree

在前期研究基础上,利用皖北地区砂姜黑土的193个土壤样本的可见近红外高光谱(350~1700nm)数据,结合非线性和线性的核函数,对9种算法进行模型的首次优化;再利用随机森林、提升树和梯度提升树三种集成学习算法进行模型组合和二次优化。通过模型比较,优选并组合了Sigmoid函数的偏最小二乘、线性的支持向量回归、径向基的支持向量回归和Sigmoid函数的支持向量回归4个单模型,集成算法优化后发现,梯度提升树算法的预测结果最优。与单模型的预测结果相比,梯度提升树模型组合的决定系数为0.86,提高了17.8%,相对分析误差系数为2.55,从B等级提升到A,不仅在准确率上有显著提高,且组合模型过拟合更低,泛化性好。因此,梯度提升树的集成学习可结合多种模型优势,通过高光谱的模型集成来提升土壤速效磷的预测结果精确度。

Based on the previous studies, visible near-infrared hyperspectral (350-1700 nm) data of 193 samples from sandy ginger black soil in northern Anhui province are firstly used to optimize the nine models by combing the nonlinear and linear kernel functions. Then, model combination and secondary optimization are performed via three integrated learning algorithms based on the random forest, boosting tree, and gradient boosting decision tree (GBDT). Four single models, including partial least squares of Sigmoid function, linear support vector regression, radial basis support vector regression, and support vector regression of Sigmoid function, are selected and combined by model comparison. After optimization of the integrated algorithms, it is found that the prediction results of the GBDT algorithm are optimal. The determination coefficient of the GBDT algorithm is 0.86, which is 17.8% higher than that of the single model, and the relative analysis error coefficient is 2.55, which is significantly improved from grade B to A. The GBDT algorithm not only improves the accuracy, but also has low overfitting degree and good generalization performance. Therefore, the GBDT algorithm can be combined with the advantages of multiple models and improve the accuracy of the prediction results of soil available phosphorus through hyperspectral model integration.