Machine Learning Frameworks for Soil Moisture Retrieval in CRNS Applications

Project Description: 

The application of cosmic-ray neutron sensing (CRNS) to environmental monitoring depends on the accurate translation of atmospheric neutron fluxes into terrestrial soil moisture values. As secondary cosmic-ray neutrons enter the Earth’s surface, they are moderated primarily by hydrogen atoms; thus, the intensity of the near-surface neutron flux is inversely proportional to the moisture content of the soil. While this relationship is well-established, the conversion process is influenced by a multitude of confounding factors, including atmospheric pressure variations, changes in absolute humidity, and the presence of hydrogen in local biomass. Traditionally, the Desilets equation—a semi-empirical analytical model—has been used to perform this calibration. However, in heterogeneous landscapes with complex vegetation or varying topographical features, these standard models often struggle to capture the non-linear dependencies and localized environmental noise that affect neutron counts over time. The accuracy of soil moisture retrieval is determined by the ability to effectively filter these external variables from the raw neutron signal. Modern computational approaches, particularly Machine Learning (ML), offer a powerful alternative to traditional calibration by identifying patterns within high-dimensional environmental datasets. By integrating neutron counts with auxiliary data—such as satellite-derived vegetation indices (NDVI), local temperature, and humidity—ML models can be trained to produce highly accurate, site-specific moisture estimates. Observations, numerical data processing, and statistical modeling are jointly used to investigate how these complex variables interact and how they can be leveraged to improve sensor autonomy. Observational data—comprising time-series neutron counts and gravimetric soil samples—provide the "ground truth" necessary for model training and validation. Numerical simulations and data-driven algorithms, such as Random Forests or Neural Networks, are used to reproduce the relationship between the neutron flux and soil moisture, accounting for environmental fluctuations. Statistical frameworks then provide the physical interpretation of these models, ensuring that the machine-learned correlations remain consistent with the underlying physics of neutron moderation and transport in the environment.(NWU)

Supervisor

North-West University (NWU)

Co-Supervisor