Accurate identification and characterization of the spatial distribution of the soil gley horizon are essential for improving soil tillage capacity and arable land quality. However, at larger spatial scales, conventional methods for gley horizon investigation—such as soil profiling and auger drilling—are constrained by low efficiency, high labor and time costs, and poor spatial continuity. Ground-penetrating radar (GPR), with its non-destructive, continuous, and large-area detection capabilities, has therefore demonstrated considerable potential for identifying soil gley horizons. Nevertheless, rapid and accurate identification of gley horizons using GPR remains challenging due to the pronounced spatial heterogeneity of soil structures and the complexity of GPR signal processing. In this study, typical gleyed paddy soils in Chengmai County, Hainan Province, China, were selected as the research object. GPR surveys using two antenna frequencies were conducted to detect the location and thickness of gley horizons in paddy soils exhibiting both surface-gley and bottom-gley characteristics. By comparing wavelet transform and variational mode decomposition (VMD) for GPR signal processing, it was found that VMD provides a stronger capability for revealing soil structural features. VMD was therefore applied to identify the vertical distribution of gley horizons, and the accuracy of gley horizon detection using the two GPR frequencies was systematically evaluated.The results indicate that, compared with wavelet transform, VMD exhibits stronger spectral amplitude intensity in the depth range of 7–36 cm, demonstrating a superior ability to characterize soil structure. In fields CM02 and CM04, the errors in gley horizon thickness detected using the 1 GHz antenna (1.5 cm and 2.0 cm, respectively) were smaller than those obtained using the 700 MHz antenna (2.4 cm and 4.0 cm, respectively), indicating higher identification accuracy at the higher frequency. The spatial distributions of gley horizons differed markedly between fields with typical surface-gley and bottom-gley features, while the coefficient of variation for gley horizon thickness across all experimental fields remained below 3%. For gley horizon depth estimation, the coefficient of determination (R²) was 0.83 at 700 MHz, with absolute errors mainly ranging from 2 to 9 cm, whereas the R² increased to 0.90 at 1 GHz, with absolute errors primarily concentrated between 3 and 5 cm. Overall, this study demonstrates that GPR combined with VMD provides an efficient and accurate approach for the rapid identification of gley horizons in paddy soils, offering strong technical support for diagnosing soil constraints and improving cultivated land quality.