SEDeepLabv3+: A New Hybrid Deep Learning Tool to Invert the High Resolution Ground Penetrating Radar Data

Least-squares optimisation problems in highdimensional parameter spaces are often used to express the difficulty of inverse issues. The deterministic gradient-based techniques that are the foundation of the current approaches have several limitations due to the non-uniqueness and linearity of the inverse issue. Here, a novel Hybrid Deep Learning Convolutional Neural Network (HDLCNN) that combines DeepLabv3+ with the Squeeze-and-excitation (Se) block, known as the SeDeepLabv3+ tool, is created to quickly forecast the subsurface picture from Ground Penetrating Radar (GPR) data. This technology has been utilised to uncover a hidden connection between the output of EM velocity and the input of GPR data. The SeDeepLabv3+ network was trained using a large number of synthetic EM velocity models and their corresponding GPR traces, and training and validation results showed $98 \%$ and $96 \%$ accuracy, respectively. The SeDeeplabv3+ inversion performs better than the Deeplabv3+ inversion in model misfit, especially in the case of GPR data misfit, and the latter is more resilient and may provide findings that are more dependable when dealing with noisy measurements. The suggested HDLCNN outperforms the conventional deterministic inversion approach in terms of obtaining resistivity data and defining the resistivity distribution. Also, compared to the commonly-used smooth velocity, the fine velocity models produced by the suggested innovative technique may better define the realworld subsurface structure, which enables the HDLCNN inversion for complicated subsurface structures whose velocities seem to vary. This way of building models might be applied to other geophysical methods now in use, making it simpler to grasp and enabling rapid and accurate calculation of the subsurface, which is a very difficult challenge.