基于随机森林与神经网络的高光谱土壤重金属Zn含量反演

在河北省保定市白洋淀区域采集115个土壤样品进行重金属含量分析和室内光谱测量，分别将BP神经网络、随机森林、决策树、多元线性回归、K近邻回归、AdaBoost回归和偏最小二乘回归法应用于全部原始波谱数据和基于双层随机森林选择后的波段数据。结果表明，基于原始波谱数据的土壤重金属Zn元素含量的反演模型精度较低，而通过双层随机森林选择出光谱数据中与土壤重金属Zn信息相关的波段，减轻了网络模型的过拟合问题，提高了模型预测精度；与其他模型比较，结合双层随机森林和BP神经网络构建的反演模型对研究区土壤重金属Zn含量预测效果最佳。

Hyperspectral Inversion of Heavy Metal Zn Content in Soil Based on Random Forest and Neural Network

A total of 115 soil samples were collected from Baiyangdian District of Baoding, Hebei Province for heavy metal content analysis and indoor spectral measurement. BP neural network, random forest, decision tree, multiple linear regression, K-nearest neighbor regression, AdaBoost regression and partial least squares regression were applied to all original spectral data and band data selected based on double-layer random forest. The results showed that the inversion model of the content of heavy metal Zn in soil based on original spectral data had a low accuracy, and the bands related to the information of heavy metal Zn in soil were selected from spectral data by double-layer random forest, which alleviated the overfitting problem of the network model and improved the model prediction accuracy. Compared with other models, the inversion model constructed by combining double-layer random forest and BP neural network had the best prediction effect on the content of heavy metal Zn in soil of the study area.