Mapping near-surface soil moisture in a Mediterranean agroforestry ecosystem using Cosmic-Ray Neutron Probe and Sentinel-1 Data

Accurate near-surface soil moisture (θ; ~ 5 cm) estimation is one of the most crucial challenges in agricultural management and hydrological studies. This study aims to map θ at high spatiotemporal resolution (17 m grid size, satellite overpass of 6 days) in a small-scale agroforestry experimental site (~ 30 ha) in southern Italy. The observation period is from November 2018 until March 2019. We employed an ensemble machine-learning method based on Random Forest (RF) to map θ. This RF method is based on three input data types: i) Sentinel-1 (S1) Synthetic Aperture Radar (SAR) measurements, ii) terrain features, and iii) supporting values of sparse point-scale θ simulated in HYDRUS-1D. We propose two different approaches to obtain supporting θ simulations via inverse modeling in HYDRUS-1D. The first approach is based on θ simulated in HYDRUS-1D, which was calibrated on soil moisture data monitored at two soil depths of 15 cm and 30 cm over 20 positions belonging to the SoilNet wireless sensor network installed in the experimental site. The second approach is based on the downscaling of field-scale θ simulated in HYDRUS-1D which was calibrated on Cosmic-Ray Neutron Probe (CRNP) data. The field-scale θ was downscaled in order to obtain sparse point-scale supporting θ over the same 20 positions by using the physical-empirical Equilibrium Moisture from Topography (EMT) model. The CRNP-based approach performed similarly to the one based on SoilNet data. Therefore, this study highlights the enormous potential for modeling reliable θ maps by integrating soft data such as S1 SAR-based measurements, topographic information, and CRNP data, having the advantage of being non-invasive and easy to maintain.