LAUSR

Portable X-ray fluorescence (pXRF) spectrometry has been successfully used for soil attribute prediction. However, recent studies have shown that accurate predictions may vary according to soil type and environmental conditions, motivating investigations in different biomes. Hence, this work attempted to accurately predict soil pH, sum of bases (SB), cation exchange capacity (CEC) at pH 7.0 and base saturation (BS) using pXRF-obtained data with high variability and robust prediction models in the Brazilian Coastal Plains biome. A total of 285 soil samples were collected to generate prediction models for A (n = 123), B (n = 162) and A+B (n = 285) horizons through stepwise multiple linear regression, support vector machine with linear kernel (SVM) and random forest. Data were divided into calibration (75%) and validation (25%) sets. Accuracy of the predictions was assessed by coefficient of determination (R2), root mean square error (RMSE), mean absolute error (MAE) and residual prediction deviation (RPD). The A+B horizons dataset had optimal performance, especially for SB predictions using SVM, achieving R2 = 0.82, RMSE = 1.02 cmolc dm–3, MAE = 1.17 cmolc dm–3 and RPD = 2.33. The most important predictor variable was Ca. Predictions using pXRF data were accurate especially for SB. Limitations of the predictions caused by soil classes and environmental conditions should be further investigated in other regions.