Long Short-Term Memory model to predict root zone soil water content from neutron count measured by Cosmic Ray Neutron Sensing

Accurate estimation of soil water content in the root zone (e.g., 0 – 30 cm) is essential for designing irrigation schedules and requires measurements that represent the field scale. Cosmic Ray Neutron Sensing (CRNS) offers a non-invasive solution that provides integrated soil moisture measurements with a horizontal footprint of approximately 7 to14 hectares and depths ranging from 15 to 83 cm, making it suitable in an area with a homogenous land use, like agricultural fields. However, CRNS sensitivity varies with both distance and depth relative to the sensor, complicating its use for estimating soil moisture in specific layers. When soil moisture is known, it is feasible to perform a forward calculation to derive neutron counts from soil water content. In this study, such calculations were performed using COSMIC, integrated with the HYDRUS-1D model. However, backward calculations, deriving soil water content from neutron counts, are not straightforward. This is because wetting and drying processes start at the soil surface, where CRNS is most sensitive. Consequently, the integrated measurement disproportionately reflects changes in the upper layers, creating a non-unique or hysteretic relationship between neutron counts and soil moisture during wetting and drying cycles. This makes predicting the 0 – 30 cm water content from neutron counts particularly challenging.

To address these limitations, we explore the application of the Long Short-Term Memory (LSTM) model to predict the average soil water content in the 0 – 30 cm layer by training the model using time series of average 0 – 30 cm soil water content and neutron counts (simulated with HYDRUS-1D COSMIC) as well as meteorological data (precipitation and reference evapotranspiration). The LSTM model is well-suited because it can learn temporal dependencies and patterns of long sequence data. The initial simulations were based on three years record of synthetic data under bare soil conditions for a region in Flanders, Belgium. While initial findings indicate a potential, further research will focus on improving the model’s robustness by training the model with more diverse variables, expanding the dataset, and integrating field measurement soil moisture records to enhance its applicability across different scenarios. This research highlights the feasibility of combining CRNS measurement, physically based modelling, and data-driven techniques to improve soil moisture estimation for irrigation management.