基于机器学习集成算法的电离层层析算法迭代初值精化

Iterative initial value refinement of ionospheric tomography algorithm based on machine learning ensemble algorithm

GNSS电离层层析技术在电离层探测中扮演着重要角色,然而,由于观测数据不足或分布不均,导致层析模型中固有的不适定性问题成为制约该技术推广应用的主要瓶颈,其主要表现在无信号射线像素的反演电子密度对初值比较依赖,而初值通常是由精度不高的经验模型给出,从而拉低了电离层层析的整体精度.针对该问题,本文提出了一种基于机器学习集成算法(XGBoost)的电离层层析新方法(XGB-CIT),即基于传统层析算法,利用多个连续时段中有信号射线像素的特征及其电子密度反演值对机器学习模型进行训练,然后预测后续时段中所有像素的电子密度,并以此作为电离层层析的迭代初值,实现对电离层层析算法迭代初值的精化以及层析精度和效率的提升.利用湖南省连续两天共46个时段的CORS观测数据进行层析反演,并以23个连续时段为滑动窗口构建机器学习模型进行初值预测,并在此基础上利用正演误差和电离层测高仪数据对XGB-CIT的精度和适用性进行检验.其结果表明:相较于IRI2016模型,XGBoost算法提供的迭代初值精度提高了68%,而基于该初值得到的XGB-CIT模型,其精度和效率也比传统层析方法有所提高,其中收敛速度提高了20%,同时根据与测高仪数据的对比分析可知,XGB-CIT对无信号射线像素的电子密度反演结果与实测数据更加吻合.

The GNSS ionospheric tomography technology plays a very important role in ionospheric detection,however,due to the insufficient or uneven distribution of observation data,the inherent ill-posed problem in the tomography model has become the main bottleneck restricting popularization and application of this technology,it is mainly shown that the inversion electron density of pixels without signal ray is dependent on the initial value,which is usually given by the empirical model with low accuracy,thus reducing the overall accuracy of ionospheric tomography.To solve this problem,this paper proposes a new ionospheric tomography method(XGB-CIT)based on the machine learning ensemble algorithm(XGBoost),that is,based on the traditional tomography algorithm,the machine learning model is trained by using the characteristics and inversion electron density of these pixels with signal ray in multiple continuous periods.Then,the electron density of all pixels in the following period is predicted and used as the initial iterative value of ionospheric tomography algorithm,which can realize the refinement of the initial iterative value that improve the accuracy and efficiency of tomography.The ionospheric inversion is carried out by using the CORS(Continuously Operating Reference Stations)observation data of 46 periods in Hunan Province for two consecutive days,and the machine learning model is constructed with 23 consecutive periods as the sliding window for initial value prediction.On this basis,the accuracy and applicability of XGB-CIT are validated by using forward modeling error and ionosonde data.The results show that the accuracy of the iterative initial value provided by XGBoost is improved by 68%compared with IRI2016,while the accuracy and efficiency of XGB-CIT model based on this initial value are also improved compared with traditional tomography method,in which the convergence speed is improved by 20%.Simultaneously,according to the comparative analysis with ionosonde data,the inversion electron density of pixels without signal ray by XGB-CIT are more consistent with the measured data.