Mois-EfficientNet: deep-learning-based 3D mapping of root zone soil moisture from GPR data

Abstract In complex soil environments, the traditional prediction methods of soil moisture often struggle to provide comprehensive spatial and temporal assessments of root zone soil moisture (RZSM). In this paper, a novel convolutional neural network, Mois-EfficientNet, is proposed based on EfficientNet-B0 for predicting average soil moisture using ground penetrating radar (GPR) data-derived images. The Mois-EfficientNet consists of two main stages: classification and regression. First, the Mois-EfficientNet is designed to classify GPR images of tree roots into distinct moisture content levels. Next, the pretrained classification network is used for transfer learning to continuously predict soil moisture content through regression tasks. The adaptive inverse distance weighted interpolation method is used to reconstruct the soil water storage distributions at various depths, facilitating the 3D noninvasive mapping of RZSM. Finally, this proposed method for mapping RZSM is validated using synthetic and field data. Compared to the actual soil moisture content of the synthetic model and that obtained from the field experiment, the prediction error follows the normal distribution. The average root-mean-square error of the moisture content derived from the proposed method is <0.02 m³ m⁻³. This new method of mapping soil moisture content in the root zone paves the way for improving water management and sustainability in the root zone.