The 1D inversion method is widely used in airborne transient electromagnetic method (ATEM) but often encounter inaccurate imaging and overlook important geological information. The three-dimensional inversion is relatively robust and provide more detailed information on subsurface resistivity distributions but limited by significant computational burdens. In this study, we develop a rapid 3D imaging scheme based on deep learning for subsurface resistivity structures with lateral heterogeneity, which performs better than traditional 1D inversion but slightly inferior to 3D inversion, but its efficiency makes it more practical than traditional 3D inversion. To create a reasonably large training set of different resistivity models that generalizes well. Here we propose three strategies that focus on training datasets, to improve the performance of DL models, including divide-and-conquer strategy, random models generating and depth constraints. Through a large of reasonable structural models, appropriate network setups, a generalized result can be obtained through our proposed UNet framework, which has been demonstrated to be effective on both synthetic and field data. The results of gradient-based 3D and 1D inversion method are compared and analyzed, respectively, which demonstrates the reliability of the proposed method. To flexibly apply the deep learning model to field data, we propose a model scaling scheme to make pre-trained UNet model compatible with different inversion specifications and survey configurations, without retraining a new DL model. The inverted structures from field data clearly delineate the geometries of the lakes, faults and surrounding mountains. The inversion operator can support instantaneous 3D subsurface resistivity imaging for ATEM observations.
