A Category-Driven Depth-Scalable Inversion Framework for 3-D GPR Data Inversion

Inversion of three-dimensional Ground-Penetrating Radar (3D GPR) data risks nonlinear electromagnetic wave propagation and the structural diversity of subsurface targets. Existing deep learning methods typically adopt fixed-depth architectures for all inputs, disregarding variations in structural complexity among varying targets. This results in redundant computation in non-target background and limited accuracy for specific types of underground targets. In response, we propose the Category-driven Depth-Scalable Inversion (CDSI) framework that integrates target classification with category-driven inversion. CDSI simultaneously extracts dynamic features (i.e., the wave changing information) via specific reservoir computing and captures spatial features through a convolutional network. A cross-attention mechanism fuses these features to predict the target category. Guided by the predicted category, a depth-scalable inversion network further adjusts its depth to reconstruct permittivity maps. Experimental results on synthetic 3D GPR datasets demonstrate that CDSI consistently outperforms existing methods in both accuracy and computational efficiency.