Research On Airborne Electromagnetic Detection Imaging Method Based On Deep Learning

Airborne electromagnetic(AEM)surveying is a emerging geophysical exploration technique in recent years.Due to its advantages of adapting to various terrains,covering large areas,and providing large amounts of electromagnetic data,it can be applied in scenarios that require fast collection of large amounts of data.It is commonly used in exploring deep geological structures,mineral resources,and environmental engineering research.However,traditional AEM inversion techniques are often time-consuming and difficult to remove noise in the late signals.To fully exploit the large volume of data generated by airborne electromagnetic detection,this thesis proposes a multi-task learning network architecture that leverages the advantages of the Transformer model in processing airborne electromagnetic sequence data,incorporates wide and deep networks to uncover the complex nonlinear relationships in the data while retaining the original sequence information.This method can overcome the drawbacks of slow inversion imaging speed and low resolution in traditional inversion techniques,as well as reduce the impact of noise in airborne electromagnetic data on imaging,achieving efficient processing and accurate imaging of the data.The specific research content is as follows:First,time-domain airborne electromagnetic forward modeling and response analysis.The forward numerical simulation of airborne electromagnetic detection is carried out using the forward response formula of magnetic dipole sources in a uniform layered geological model.The simulated results are used to analyze the data characteristics of timedomain airborne electromagnetic and to investigate factors that affect the response of airborne electromagnetic exploration,including flight height and the magnetic moment of the transmitter coil.Second,comparative study of time-domain airborne electromagnetic imaging methods based on deep learning.Different geological resistivity models are established,and a sample set of AEM data is constructed.A neural network model for imaging in AEM exploration is established,and different network structures(fully connected network,convolutional network,sequence-to-sequence network,and transformer network)are studied for their effectiveness in AEM data inversion imaging.Suitable hyperparameters and optimization methods are tuned,and the optimal network is determined to be the transformer combined with the wide and deep network structure.The imaging performance is validated on uniform half-space and layered typical geological models,and the network is able to image the trend of geological resistivity changes.Third,neural network processing method for airborne electromagnetic data with noise.By optimizing the Transformer model combined with a wide and deep network structure and incorporating multi-task learning,a sub-network for simultaneous denoising and imaging is established,with the imaging and denoising branches mutually constrained.Tests with a noise test set were conducted to validate the anti-interference ability of the imaging branch against noise,which improved imaging accuracy.The denoising branch effectively improved the signal-to-noise ratio,with an average improvement of 13.98 dB over the entire noise test set.Fourth,testing the effectiveness of the proposed multi-task network model on field exploration data.The multi-task learning network structure,which combines Transformer and wide and deep networks,is proposed and validated to accurately predict single points,profiles,and cross-sections in the exploration area,in agreement with the exploration results.The imaging performance of the multi-task model is validated to be superior to Seq2Seq and Transformer methods in terms of accuracy.The algorithm’s anti-interference ability to noise is also verified,as it can extract geological resistivity information from profiles with locally high noise levels.