GPR

GPR is an acronym with multiple meanings used in different fields. The two most common in science and engineering are ground-penetrating radar, a geophysical imaging method, and Gaussian process regression, a Bayesian non-parametric approach to regression in machine learning. In computing, GPR can also refer to a general-purpose register, a CPU register used for general arithmetic and data manipulation.

Ground-penetrating radar involves transmitting short pulses of high-frequency electromagnetic energy into the ground and recording the reflections from subsurface interfaces. A GPR system typically uses an antenna carried along a survey line, producing a two-dimensional profile or three-dimensional data when deployed in grid patterns. By analyzing travel times and amplitudes of reflections, practitioners infer subsurface features such as buried utilities, voids, stratigraphy, or buried archaeological remains. Frequency ranges commonly span from around 10 to a few hundred megahertz, with higher frequencies providing finer resolution but shallower penetration. Limitations include variability in soil properties, signal attenuation, and clutter from conductive materials; accurate interpretation often requires calibration, velocity models, and corroborating data.

Gaussian process regression treats unknown functions as random draws from a Gaussian process, defined by a mean function and a covariance kernel. Given observed data, it yields a posterior distribution over functions and predictive uncertainty for new inputs. Common kernels include squared exponential and Matérn families, with hyperparameters learned from data. GPR is flexible for small to moderate data sets and provides uncertainty estimates, but it has computational complexity O(n^3) for n training points and can be sensitive to kernel choice. It is widely used for regression, kriging-like spatial modeling, surrogate modeling, and Bayesian optimization.