Mapping 3D Root Zone Soil Moisture of GPR Data Based on Deep Learning

Traditional methods for soil moisture prediction often face challenges in providing comprehensive spatial and temporal assessments of root zone soil moisture (RZSM) in complex soil environments. This study proposes a novel approach based on the convolutional neural network (CNN) for predicting average soil moisture based on images obtained from ground penetrating radar (GPR) data. The CNN is structured in two main stages: classification and regression. First, the CNN classifies GPR images of tree roots into distinct moisture content categories. Then, the pre-trained classification network is adapted using transfer learning to perform regression tasks, predicting continuous soil moisture values. To enable 3D non-invasive mapping of RZSM, we apply adaptive inverse distance weighted interpolation to reconstruct the distribution of soil water storage at various depths, ultimately generating a 3D visualization of RZSM. Finally, we validate the proposed approach using both synthetic and field data of GPR. The root mean square error between the soil moisture content predicted by this approach and the actual moisture content of the synthetic model, as well as the moisture content obtained in a field experiment, is less than 0.02 m³·m⁻³. This new approach for mapping RZSM holds great potential for enhancing root zone water management and promoting sustainability.