结合GA-BP神经网络的GPS-IR雪深反演

GPS-IR snow depth retrieval combined with GA-BP neural network

针对传统雪深反演中出现的系统偏差和跳变问题,该文提出一种结合遗传算法-反向传播(GA-BP)神经网络的雪深反演方法。首先通过二次项拟合有效分离出信噪比残差,进而对变换单位后的信噪比残差进行频谱分析,计算得到初步雪深值。最后,建立基于初步雪深值的GA-BP神经网络优化模型。以美国板块边界观测计划(PBO)提供的监测数据为例,并与传统方法对比分析,结果表明:采用GA-BP神经网络不仅能够削弱初步反演结果中出现的系统偏差,还能有效消除反演过程出现的跳变现象。采用PRN09和PRN24卫星反演,RMSE和MAE均分别小于0.083 m和0.065 m,R^2有了明显提高,优于未处理初始雪深的情况。

Aiming at the problem of system deviation and cycle slips in traditional snow depth retrieval,a method of snow depth retrieval combining genetic algorithms-back propagation(GA-BP)neural network was proposed in this paper.Firstly,the residual of signal to noise ratio(SNR)was separated effectively by quadratic item fitting method from original sequence,and then the spectrum analysis for the SNR residual which had been transformed unit was carried out,and the initial snow depth was calculated,and finally,a GA-BP neural network optimization model based on the initial snow depth value was established.Taking the monitoring data provided by plate boundary observatory(PBO)as an example,and compared it with traditional methods.The results showed that using GA-BP neural network could not only weaken the systematic bias in the preliminary inversion results,but also effectively eliminate the cycle slips phenomenon in the inversion process.Using the PRN09 and PRN24 satellites for inversion,the root mean square error(RMSE)and mean absolute error(MAE)were less than 0.083 mand 0.065 mrespectively,and R^2 was significantly improved,which was superior to the untreated case.