Many ground-penetrating radar (GPR) studies incorporate tomographic methods that use straight raypaths for direct model reconstruction, which is unrealistic for media with gradually changing petrophysics. Ray-bending algorithms can sometimes lead to unreliable resolution, especially at interfaces of abrupt dielectric changes. We present an improved GPR tomography technique based on a combination of seismic tomographic methods and a finite-difference solution of the eikonal equation. Our inversion algorithm uses velocity gradient zones and bending rays that represent realistic geology in the subsurface. We tested the technique on theoretical and experimental models with anomalous bodies of varying saturations and velocity and applied it to data from a GPR field experiment thatanalyzed the root zones of trees. Synthetic results showed that the resolution of our technique is better than that of published methods, especially for local anomalies with sharp velocity contacts. Our laboratory experiments consisted of four objects buried in sand with various water saturations. The GPR tomogram could map the objects and determine their degree of saturation. The velocities are compatible with those of the complex refraction index method; their relationship to the water content fits a previously published empirical equation. Our original field experiment around a poplar tree could map the heterogeneous subsurface and distinguish a central low velocity beneath the tree from the peripheral negative anomaly of a refill. This zone reflects the whole root zone and is caused by its bulk water content of both the organic root network and its surrounding soils.