AI‐Powered GPR Characterization of Peatland Underground Cavity Induced by Smoldering Wildfire: A Laboratory Proof‐of‐Concept

ABSTRACT

Global peatlands are increasingly vulnerable to severe and persistent wildfires under climate change. Once ignited, peat fires often smolder beneath the surface, forming underground cavities that are challenging to detect using conventional sensing techniques. This work proposes a novel approach that integrates ground‐penetrating radar (GPR) with artificial intelligence (AI) to estimate the geometry of idealized air‐filled cavity analogs, providing a support basis for characterizing and quantifying the fire‐induced underground cavities. Laboratory experiments were conducted for spherical cavities created by embedded balloons in a 30 cm‐thick peat soil layer, and the underground cavity has a depth ranging from 0.5 to 15 cm and a diameter varying from 3 to 20 cm. Experiments showed that GPR imaging successfully captured reflection patterns associated with these ideal cavity boundaries in regular geometry, and data were subsequently trained by a ResNet‐based deep learning model. The model achieved strong preliminary evidence, with total coefficients exceeding 0.89, as well as acceptable values for cavity diameter, minimum depth and center depth, respectively. Further analysis showed that incorporating multiple image filters and repeated measurements led to modest improvements in prediction accuracy, whereas additional feature extraction provided limited benefit. This research provides a laboratory proof‐of‐concept for noninvasive detection of underground smoldering‐induced cavities in peatlands, offering the first step of valuable support for assessing fire severity, evaluating ecological damages, and informing wildfire management strategies.