Geostatistics and Spatial Modelling in Soil Science: A Critical Appraisal of Methods, Advantages, Inconveniences and Applications

Abstract

Soil properties vary continuously across space and time, making spatial prediction at unsampled locations fundamental to soil science. Geostatistics has provided a coherent framework for quantifying, mapping and communicating this variation for several decades. However, its role is increasingly challenged by machine learning, extensive environmental covariates from remote and proximal sensing, and growing demands for national and global soil information with reliable uncertainty estimates. This review critically evaluates the methodological foundations, strengths, limitations and applications of geostatistical and hybrid spatial modelling in soil science. Relevant literature was identified through structured searches of scholarly databases and citation tracking and assessed for methodological adequacy, validation quality, transparency and relevance. Five major issues emerge. First, key assumptions underlying kriging, particularly stationarity of the mean and covariance structure, are frequently violated in soil datasets, while the implications are often inadequately reported. Second, variogram estimation remains central to geostatistics, but limited sample sizes in many studies reduce its reliability and weaken interpretations of spatial dependence. Third, comparisons among kriging methods, regression-kriging and machine learning show inconsistent performance, largely because of differences in sampling design, covariate quality, extrapolation requirements and validation procedures. Fourth, although uncertainty quantification has improved conceptually, its practical validation remains inadequate, raising concerns about reported model accuracy. Fifth, applications in precision agriculture, contamination assessment, salinity management and soil carbon accounting differ substantially in evidential maturity, with soil carbon monitoring showing notable gaps between mapped precision and decision-grade reliability. Overall, geostatistics remains important for representing spatial dependence, designing sampling strategies and quantifying uncertainty. Future advances are likely to be strongest when geostatistical methods and machine learning are integrated rather than treated as competing alternatives.