Abstract
Hyper-arid agricultural soils combine low organic carbon, high calcium carbonate contents, salinity, and frequent gypsum accumulation, making rapid quantification of soil organic carbon (SOC), soil inorganic carbon (SIC), and gypsum valuable but analytically demanding. We evaluated visible–near infrared (VNIR) spectroscopy for predicting these three properties in 216 samples from 18 plots, nested within six oasis sites spanning two oasis management systems, three topographic positions, and six soil depths, with particular emphasis on how validation design affects apparent model accuracy. Partial least squares regression (PLSR) and random forest (RF), single- and multi-target workflows, and global versus stratified calibrations were compared using plot-blocked GroupKFold, site-blocked GroupKFold, random KFold, and leave-one-out cross-validation (LOOCV). Under continuum-removed spectra, RF outperformed PLSR across all three targets. Predictability was strongly target-specific: gypsum was predicted most reliably (R² = 0.78), SOC moderately (R² = 0.44), and SIC least reliably (R² = 0.30). KFold and LOOCV generally yielded more optimistic estimates than plot-blocked GroupKFold, inflating SIC R² by 0.15–0.19 for the global model and by up to 0.64 for stratified models. Site-blocked validation further showed that spatial transferability was strongly target-specific: SOC performance remained similar between plot-blocked and site-blocked validation (R² = 0.41 and 0.43, respectively), gypsum performance declined moderately (R² = 0.78 to 0.64), whereas SIC performance decreased markedly from R² = 0.30 to −0.28, indicating poorer performance than a mean-based prediction under oasis-level blocking. VNIR spectroscopy therefore shows strong potential for gypsum, moderate utility for SOC, but limited transferability for SIC. Validation design, together with the spatial scale at which independence is imposed, critically determines how soil-spectroscopy performance should be interpreted for dryland monitoring.