Soil health assessment in a dryland ecosystem: mapping soil properties with PRISMA and EnMAP hyperspectral data and machine learning

Quantifying soil health (SH) in drylands has become a global priority, yet conventional soil surveys remain too costly and sparse to capture its high spatial heterogeneity. Next-generation spaceborne hyperspectral sensors offer new opportunities for regional soil monitoring. However, it remains unclear whether hyperspectral imagery alone can provide a reliable remote sensing–based proxy of SH without ancillary environmental covariates. We evaluated PRISMA and EnMAP hyperspectral imagery coupled with machine learning (ML) to predict soil properties and develop a soil health index (SHI) for Sirjan Playa, Iran. Using 244 bare-soil samples, four ML algorithms, including partial least squares regression (PLSR), Gaussian process regression (GPR), support vector regression (SVR), and random forest (RF), were evaluated under 10-fold cross-validation to predict salinity, pH, calcium carbonate equivalent (CCE), gypsum, and soil texture, which are relevant indicators of SH in drylands. The predicted soil property maps were integrated into the SHI using weights derived from the analytic hierarchy process (AHP). Our findings highlight the reliable performance of both sensors for most soil properties. Gypsum and CCE achieved the highest ratio of performance to interquartile range (RPIQ) values, with 3.20 and 2.90 for PRISMA and 4.86 and 3.32 for EnMAP, respectively. Salinity was also reliably predicted (RPIQ = 2.40 for PRISMA and 2.10 for EnMAP), although uncertainty was higher in areas with extremely high salinity and scarce soil observations. The resulting SHI was consistent with vegetation performance, as reflected in the 2019–2024 Sentinel-2 normalized difference vegetation index (NDVI) time series, with higher NDVI values observed in high-SHI zones than in constrained areas (up to 0.57 vs. <0.09), demonstrating consistency between SH conditions and plant vigor. Sirjan Playa was characterized by predominantly medium-to-low SHI, identifying salinization as the main driver of soil degradation. Overall, integrating hyperspectral imagery and ML for SHI mapping provides insights into soil degradation and supports evidence-based land management and policymaking in dryland ecosystems.