变分模态分解和机器学习融合的GNSS-IR土壤湿度反演

GNSS-IR soil moisture inversion combined with variational mode decomposition and machine learning fusion

针对如何有效去除GNSS-IR土壤湿度反演中卫星信号噪声比例高、地表粗糙度带来的散射影响等问题,建立了一种将变分模态分解与BP神经网络相结合的模型,该模型利用变分模态分解自适应与非递归的特点来替代传统的多项式拟合法,从而有效提高反射信号提取精度;并利用BP神经网络的非线性映射能力进行后期预测,与传统线性回归进行对比分析。利用PBO H_(2)O的土壤湿度作为参考依据,以2016年PBO 783测站的GNSS数据为基础建立模型并评估分析。实验结果表明:结合变分模态分解与BP神经网络的土壤湿度模型反演结果与参考数据在大体趋势上基本一致,其均方根误差为0.014,决定系数R^(2)为0.951,对比单星线性回归模型提升了42.79%,证明了该方法确实能够有效提高反射信号质量及抑制地表粗糙度影响从而提高土壤湿度反演精度。

In order to effectively remove the proportion of satellite signal noise and the scattering influence caused by surface roughness in GNSS-IR soil moisture inversion,a model combining variational modal decomposition and BP neural network is established.This model uses the adaptive and non-recursive characteristics of variational modal decomposition to replace the traditional polynomial fitting method,thereby effectively improving the extraction accuracy of reflected signals;The nonlinear mapping ability of BP neural network is used for later inversion,which is compared with the traditional linear regression.Using the soil moisture of PBO H_(2)O as the reference basis,the model is established and evaluated based on the GNSS data of PBO 783station in 2016.The experimental results show that the inversion results of soil moisture model combined with variational modal decomposition and BP neural network are basically consistent with the measured data,the root mean square error is 0.014,the determination coefficient R^(2)is 0.951,which is 42.79%higher than that of single satellite linear regression model,It is proved that this method can improve the quality of reflected signal and restrain the influence of surface roughness,so as to improve the accuracy of soil moisture inversion.