Optimal Calibration Sizes for Predicting Soil Properties Using Vis‐ <scp>NIR</scp> Spectra Across Scales

ABSTRACT Soil visible near‐infrared (vis‐NIR) spectroscopy has demonstrated significant potential in providing accurate soil information in a cost‐effective manner, which is crucial for providing updated soil information. However, building a comprehensive soil spectral library requires substantial financial resources, making it essential to balance cost and accuracy for soil spectroscopic predictions. Despite its importance, there has been no systematic comparison of how soil properties, calibration models, and spatial scales impact the optimal calibration size for these predictions using a consistent soil spectral library. This study addresses this gap by utilizing LUCAS Soil 2009 data to determine the optimal calibration size for soil organic carbon (SOC), pH, clay, and cation exchange capacity (CEC) using Partial Least Squares Regression (PLSR), Cubist and Random Forest (RF), Convolutional Neural Network (CNN), and Memory‐Based Learning (MBL) algorithms at regional, national, and continental scales. Our findings indicate that MBL and Cubist consistently outperformed other algorithms across all scales, particularly at larger spatial scales, while CNN showed comparable performance at national and continental scales when using large calibration sample sizes. The optimal calibration size (the minimal number of calibration samples that reach the plateau of model performance) was influenced by the specific soil property, calibration algorithm, and scale: (1) plateau calibration sizes were identified as 150–250, 600–1000, and 2000–6000 samples for regional, national and continental scales, respectively; (2) SOC, clay and CEC required more calibration samples than pH at national and continental scales; (3) CNN required more samples for superior performance than MBL and Cubist at larger extents (national to continental) while it has the potential to perform better with larger calibration size. These outcomes offer valuable guidance for the cost‐effective design of soil sampling strategies that support sustainable soil management across different spatial scales.