Performance of interpolation methods in digital soil mapping: the influence of data characteristics

Purpose The selection of interpolation methods in digital soil mapping lacks a systematic approach, reducing map accuracy. This study aimed to evaluate whether data characteristics, such as sample size and spatial structure, influence the selection and performance of interpolation methods. Methods Six interpolation methods were evaluated across datasets representing three typical sampling density scenarios in Brazilian agriculture. Spatial structure was characterized using Moran’s index and the spatial dependence index derived from geostatistical semivariograms. Interpolation was performed, and the accuracy was assessed using test datasets and Lin’s concordance correlation coefficient. Consequently, two decision frameworks (multivariate and univariate) were developed to guide method selection. The univariate framework was then validated to assess its robustness. Results For small datasets (n < 50), deterministic methods, particularly Thin Plate Spline (TPS), consistently provided the most stable predictions. In contrast, the performance of the geostatistical and machine learning methods improved with increasing sample size and stronger spatial structure. In the largest datasets (n ≥ 100), most methods became competitive, shifting the primary selection criteria towards factors such as operational simplicity. These findings were synthesized into decision frameworks to guide optimal interpolator selection. Conclusion Interpolation performance is critically dependent on underlying data attributes (sample size and spatial structure). No universal interpolator exists for all datasets. Deterministic methods, specifically the TPS, demonstrated superior flexibility across diverse scenarios. A data-driven decision framework was developed in this study translating these key data attributes into clear, actionable recommendations, thereby providing users with an accessible tool to demonstrably improve the reliability of soil maps.