Portable X-ray fluorescence spectrometry library for Brazilian soils: chemical characterization and prediction of soil particle size fractions across all biomes

Sensor-based soil data libraries have proven valuable for a wide range of applications. Given the increasing adoption of portable X-ray fluorescence (pXRF) spectroscopy for soil-related studies, we aimed to create a library of soils characterized by pXRF and develop prediction models for soil particle size fractions (clay, silt, and sand contents) based on 16,989 samples collected at multiple soil horizons, from all Brazilian biomes (Amazon, Caatinga, Cerrado, Atlantic Forest, Pantanal, and Pampa), under multiple climate zones, land uses and derived from several parent materials. Random Forest (RF), Support Vector Machine (SVM), Gradient Boosting Machine (GBM), and Genetic Programming (GP) were used to train the models. The dataset was split into training (13,593) and independent validation (3,396) sets. The accuracy of models was assessed via root mean square error (RMSE) and coefficient of determination (R 2 ) values. Sand, silt and clay contents ranged from 0 to 100%, 0 to 96%, and 0 to 93%, respectively. Strong correlations were found between particle size fractions and Fe and Si contents, but some strong correlations were biome-dependent, such as those between K and silt (0.77) in Pantanal. Random Forest outperformed the other algorithms, achieving RMSE and R 2 values, respectively, of 8.2% and 0.89 for sand, 5.6% and 0.71 for silt, and 6.3% and 0.89 for clay contents. This pXRF library resulted in adequate predictions for sand and clay contents. This national library may contribute to establish a global library of soils characterized via pXRF. This dataset should be continuously increased to support soil chemical characterization and develop prediction models for varied soil properties.