Detection of Buried Landmines using a Convolutional Autoencoder trained on Simulated prompt Gamma Spectra

Abstract

The detection of buried landmines remains a persistent challenge in security and humanitarian demining. In this work, we present an indirect detection methodology based on the analysis of prompt gamma-ray emissions induced by 14 MeV neutron irradiation. A high-resolution LaBr₃ detector captures the gamma spectra arising from neutron interactions with soil constituents and buried explosives. A Convolutional Neural Network (CNN) autoencoder, trained in an unsupervised manner, models the intrinsic spectral response of soil under varying moisture conditions. Anomalies between reconstructed and measured spectra are used to infer the presence of subsurface anomalies consistent with landmines. Monte Carlo simulations, conducted with the Geant4 toolkit, generate a comprehensive dataset encompassing a soil matrix under various moisture levels. The proposed system demonstrates sensitivity to buried antipersonnel landmines at shallow depths, validating the integration of neutron activation analysis and deep learning for advanced landmine detection applications.