Multi-depth soil moisture estimation via 1D convolutional neural networks from drone-mounted ground penetrating Radar data

Accurate soil moisture estimation is pivotal for sustainable agricultural practices, directly influencing irrigation management, crop yield optimization, and water conservation. Precise moisture assessment ensures the efficient use of water resources, which is critical in the context of climate variability and the increasing scarcity of freshwater. However, a significant challenge in soil moisture estimation has been the limited capability of conventional remote sensing techniques to penetrate biomass and provide accurate moisture data at multiple soil depths. These methods often struggle to differentiate between surface moisture and the moisture profile across the soil column, especially under dense vegetation cover. This study introduces an advanced methodology using a 1D Convolutional Neural Network Artificial Neural Network (1D-CNN-ANN) to process raw Ground Penetrating Radar (GPR) amplitude data for enhanced soil moisture estimation. The 1D-CNN-ANN model is designed to tackle the multi-depth moisture estimation challenge and penetrate through biomass to yield accurate soil moisture readings. Our approach is validated by comparing the 1D-CNN-ANN model's performance with two other established machine learning algorithms: Gradient Boosting Machine (GBM), and Support Vector Machine (SVM). The results demonstrate the 1D-CNN-ANN model's superiority, reflected in higher R2 values and lower error metrics (RMSE) across different soil depths compared to the other machine learning models. At a depth of 10 cm, the 1D-CNN-ANN model achieved an R2 value of 0.84, outperforming the GBM, and SVM models with R2 values of 0.77 and 0.79, respectively. At 20 cm and 30 cm depths, the model also showed superior accuracy with R2 values of 0.82 and 0.74, respectively. Notably, the 1D-CNN-ANN model maintained its robust performance even at the challenging depth of 40 cm, with an R2 of 0.64, where traditional sensors typically falter due to biomass interference. Conceptually, the study reveals the model's capacity to discern the complex patterns of soil moisture across a range of conditions, offering valuable insights for irrigation scheduling and water resource management in both irrigated and non-irrigated agricultural settings.. © 2025 The Authors