Neural networks for optimal initial guess selection in nonlinear microwave subsurface imaging

Subsurface non-destructive exploration is of interest for several applications which span from archeology and civil engineering to safety and security. Among the available remote sensing imaging modalities, ground penetrating radar (GPR) seems very attractive for the detection and characterisation of buried targets. This technology exploits electromagnetic signals in the microwave band to perform the exploration in a non-invasive way. Unfortunately, conventional GPR images require an expert user to be interpreted and do not provide quantitative information about the buried objets. In this framework, this paper explores the use of artificial neural networks for quantitative multiple-input-multiple-output tomographic ground penetrating radar to improve the reliability of the quantitative tomographic recovery, speeding up the nonlinear inversion and providing a user-friendly image which is easy to be interpreted even for a non-expert user.