Neural network inversion of seismic wave velocities for vadose zone water content profile

Accurate estimation of water saturation in the vadose zone is crucial for hydrological and agricultural applications. However, traditional seismic inversion methods often struggle with non-linearity and sensitivity to noise, limiting their generalisation capabilities. To address this challenge, we propose a neural network-based inversion approach that allows the assessment of the vertical distribution of water saturation from compressional ($v_{{}_P}$) and shear ($v_{{}_S}$) wave velocities. Specifically, our architecture is based on a regression model and incorporates an autoencoder layer to improve robustness against noise. As a result, this enhances its ability to invert complete saturation profiles and increases its adaptability to complex hydrological conditions. Furthermore, the model demonstrates strong performance across varying water table depths, with low error metrics and high resilience to input noise with a RMSE of 3.34 × 10⁻² and a R${}^2$ of 0.978 for 5% noise. Our current approach has been trained exclusively on noisy synthetic data. We plan to validate it in the near future against experimental field data we have recorded for an agricultural soil. Overall, this study establishes a foundation for future applications of deep learning in hydrogeophysical inversion and underscores the need for validation with real-world data.