Time Domain Airborne Electromagnetic Imaging Based On Neural Network

The time domain airborne electromagnetic system(AEM)has the advantages of high sampling rate,rich information and high resolution.The high sampling rate results in a large amount of time-domain field data,which cannot efficiently be interpreted by two-dimensional or three-dimensional inversion.Therefore,imaging and onedimensional inversion are the main method for data processing.Since one-dimensional inversion will take a long time to run when processing a large amount of data and the result often depends on initial model,the convenient,fast and stable imaging methods are still the most popular method for processing field data.In order to overcome the shortcomings of low resolution of conventional imaging methods,we propose a new imaging method based on neural network.Based on the numerical experiment and analysis,we adopt the neural network structure of the convolutional neural network(CNN)and long short term memory(LSTM)to do the imaging.This kind of deep neural network structure can be used to learn and store the complex nonlinear relationship between resistivity models and electromagnetic(EM)signals,thus it can produce high-precision imaging results closed to one-dimensional inversion.In order to test the validity of this new method,we first test it on the synthetic datasets.The results show that this method can achieve highspeed and accurate imaging with reasonable network parameters.Since the flight altitude parameter is the very important parameter which affects the electromagnetic signal serivously,we increased the weight of the flight altitude in the training process to realize accurate imaging at different flight height.In order to deal with the noise in the field data,the early stopping method is used to train the network.Firstly,we test the network before processing the field data,and then select the optimal network parameters by qualitative and quantitative analysis.Finally,we suppress the noise and obtain the reliable imaging results.Furthermore,we study the influence of the generation method of training set and training set size on network training results,and test the performance of neural network trained by different training sets.The test results show that the training set composed of smooth underground resistivity model performs better in convergence error and speed than the regular training set.In order to verify the practicability of this method,we image the field data and compare it with the onedimensional inversion results.The results show that the imaging results by smooth model trained network are closer to the inversion results of Occam method.Through the reliability analysis of the imaging results,it is found that the neural network trained by the regular training set is better than the neural network trained by the smooth model.At last,we do some numerical experiments on the time domain airborne electromagnetic 3D imaging based on neural network.For the simple abnormal body buried in uniform halfspace,the neural network can accurately recover its position,shape and resistivity,which means that the neural network imaging method has the potential for 3D imaging.The imaging accuracy of the neural network imaging method developed in this paper is very close to that of one-dimensional inversion,but its computational efficiency is far higher than that of the conventional one-dimensional inversion,thus it is a new method with great potential.In addition,after the successful training,the neural network can image the AEM data generated by the same EM system without any additional processing.This new method has strong flexibility and portability,and has broad application prospects.