Research On The Denoising Methods Based On Gaussian Process Regression For Time Domain Airborne Electromagnetic Data

Airborne electromagnetic method in time domain is a geophysical exploration method that uses aircraft as the carrier to detect the electrical distribution of underground medium by measuring the secondary field induced by underground medium changing with time.It has the advantages of wide exploration area,relatively low cost,little influence by topography and landform,high exploration efficiency and good horizontal resolution.Therefore,it can be widely used in many fields therefore,such as: groundwater distribution survey and mineral resource exploration,detection of unexploded ordnance,geological mapping and geological hazard monitoring,etc.The time domain airborne electromagnetic method usually collects the weak secondary field response of the signal,the effective signal frequency band is wide,so that the received signal is easily affected by many kinds of noise,the exploration depth of the underground medium and the data processing and inversion imaging accuracy in the later period are restricted.Therefore,it is of great significance to improve the quality of data processing and the precision of inversion interpretation by preprocessing the time-domain airborne electromagnetic data and improving the signal-to-noise ratio of the data.Conventional noise suppression methods mostly start from a single noise and filter from the perspectives of time domain,frequency domain and signal decomposition.However,there are overlaps between airborne electromagnetic signals and noise in the time and frequency domain,and conventional treatment methods are often unable to achieve the desired effect.Since many phenomena or signals can be regarded as random signals,it is one of the important means to study signal processing from the perspective of statistics.In this paper,Gaussian Process Regression method is used to study the de-noising of airborne electromagnetic data in time domain and analyze its effectiveness and feasibility under the support of "research and development of helicopter airborne electromagnetic detection data interpretation software system",a national key research and development project of the 13 th five-year plan.(1)Generation of Time-domain airborne electromagnetic simulation data.First,the one-dimensional airborne transient electromagnetic response data based on the central loop device is obtained by using the one-dimensional forward modeling method.Then,a layered two-dimensional earth model is established on the basis of the 2.5-dimensional theory,and a 2.5-dimensional program is used to generate the simulated 2-d electromagnetic response.Finally,according to the Gaussian characteristics of noise and the characteristics of airborne noise in the measured data,Gaussian white noise and atmospheric noise are added on the basis of 1-d and 2-d electromagnetic data to obtain the simulated noise-containing data,so as to provide basic test data for the subsequent research on de-noising methods and the analysis of de-noising results.(2)The de-noising method based on Gaussian process regression model.Firstly,aiming at the selection of kernel function in Gaussian process regression,a Gaussian prior to the definition of common square exponential covariance function and rational quadratic covariance function is used for sampling experiments,the analysis of which kernel function is superior can be used as model support for subsequent de-noising experiments.Then,aiming at the problem of super-parameter selection in Gaussian process regression,the normal linear fitting experiment was conducted from a single parameter(length scale)to analyze the influence of different parameter values on the model fitting effect,so as to provide theoretical basis for the adaptive acquisition of super-parameter values in subsequent de-noising experiments.Finally,Gaussian process regression(GPR)is used to de-noising the airborne electromagnetic one-dimensional simulation data in time domain,and its validity and feasibility are analyzed.(3)The de-noising method based on sparse Gaussian process regression.Since the covariance function needs to be inversed in the Gaussian process regression method,the traditional Gaussian process regression method needs to consume a large amount of calculation cost.Therefore,this paper adopts the sparse Gaussian process regression method to reduce the operation cost and improve the prediction efficiency.Firstly,a prediction experiment is conducted to add a certain noise to the sine model to prove the feasibility of the sparse pseudo-input method in de-noising.Then,in the case of a fixed number of pseudo-input sets,curve fitting experiments are carried out based on different training points to analyze the influence of training points on prediction results and prediction speed.Finally,in the case of fixed training points for different number of pseudo input sets sparse pseudo input method and the traditional Gaussian process regression method is applied to the time-domain airborne electromagnetic one-dimensional simulation data de-noising,and through the signal-to-noise ratio and the root mean square error evaluation both de-noising effect,through the training and predicting time efficiency evaluation of two methods of de-noising.(4)Performance comparison of Gaussian process regression and sparse Gaussian process regression de-noising methods.Firstly,Gaussian process regression and sparse Gaussian process regression are used to de-noising the electromagnetic response data of the 2.5-dimensional simulated 2-d profile.Then,the de-noised data are inversed,the de-noising effect of the two methods is analyzed.Finally,the two methods are used to process the measured data in a certain exploration area in Xinjiang to further verify the feasibility of the proposed de-noising method.