Hierarchical modelling to map soil salinity using proximal sensing and a multispectral UAV

Soil salinity commonly exhibits strong within-field variability, and operational diagnosis requires scaling point-based reference measurements to spatially continuous estimates. This study evaluates a hierarchical, cascading modelling framework to map soil salinity at field scale in agricultural plots in Belorado (Burgos, Spain), integrating laboratory reference data with diffuse reflectance spectroscopy, proximal apparent electrical conductivity sensing, and very-high-resolution multispectral UAV products.

In each plot, 40 sampling locations are established and georeferenced using GNSS RTK. Soil samples collected at these locations are analysed in the laboratory to obtain salinity reference values. On the same day, and prior to the commencement of sampling operations, a multispectral UAV survey (DJI P4 Multispectral) is conducted to generate high-resolution orthomosaics and derived spectral variables. This is followed by a full-coverage proximal survey using a VERIS Q2800 system to measure apparent electrical conductivity (ECa) continuously across each plot. After these surveys, sampling is performed and diffuse reflectance spectra are acquired in situ at the GNSS-referenced locations using a NeoSpectra (Si-Ware) instrument. In addition, laboratory spectroscopy is repeated on air-dried samples to quantify moisture effects and to assess the consistency between field and laboratory spectral acquisitions.

The upscaling strategy is implemented as a cascade of transfer models: (i) laboratory salinity and diffuse reflectance spectroscopy, (ii) diffuse reflectance spectroscopy and ECa (VERIS), and (iii) ECa and UAV-derived spectral variables, enabling plot-wide prediction. Model performance is assessed using cross-validation, including spatially explicit schemes, and uncertainty propagation along the cascade is examined where feasible. The outcome is a reproducible workflow for producing field-scale salinity maps and quantifying the added value of each sensing layer in the hierarchical framework.