Impact of calibration set size for predicting soil fertility attributes using local pXRF spectral libraries

Employing portable Energy-Dispersive X-Ray Fluorescence (pXRF) sensors and Machine Learning (ML) using calibration set sizes with reduced samples can save resources and time in measuring soil fertility attributes. The calibration of local models must have a sufficient number of samples to represent the spectral variation of the area related to the properties of interest, avoiding excess, which would be redundant, and lack, which would cause a loss of performance. From this perspective, this study investigated the impact of different calibration set sizes on the performance of Partial Least Square Regression (PLS) models using a pXRF sensor (0‐15 kV) to predict Exchangeable Potassium (K⁺), and Calcium (Ca²⁺), Cation Exchange Capacity (CEC), and Soil Organic Carbon (SOC). Two Brazilian tropical fields composed the pXRF spectral libraries. The Cambé and Toledo datasets comprised 386 and 205 soil samples, respectively. Applying the Kennard-Stone sampling selection algorithm, calibration set sizes ranged from 36 to 276 samples, with increments of 20, for the Cambé field, and from 20 to 140 samples, also with increments of 20, for the Toledo field. Results evaluated on validation sets indicate that calibration sample size can be optimized without compromising model accuracy. Established by quantitatively comparing the models RMSEP, the optimal number of calibration samples was attribute- and field-specific. For the Cambé library, the optimal calibration set size was 116, 136, 156 and 116 for the SOC, CEC, K + and Ca 2+ , respectively. For the Toledo library, the SOC, CEC, K + , and Ca 2+ had optimal sample numbers of 60, 100, 80, and 80, respectively. All these models presented satisfactory RPIQ values (≥ 1.7). Therefore, to encompass all the evaluated attributes, it is suggested 156 calibration samples as a reference for developing local libraries applying pXRF sensors in tropical soils with oxide-rich mineralogy. • Calibration sample size was optimized without compromising model accuracy. • The optimal number of calibration samples was attribute- and field-specific, ranging from 60 to 156. • 156 samples are the minimum XRF library size for simultaneous prediction of soil attributes.