Soil Advances (Dec 2025)
Portable XRF and Vis-NIR spectrometry for predicting chemical properties of forest soils in the Amazon: Insights into sensor data dimensionality reduction
Abstract
Proximal sensor data integrated with machine learning algorithms have been successful in assessing soil chemical parameters. The accuracy of these sensors for assessing soil parameters in tropical forest soils is still limited. This research employed portable X-ray fluorescence spectrometry (pXRF) and visible and near-infrared spectrometry (Vis-NIR) sensors to predict soil pH, soil organic matter (SOM), and cation exchange capacity (CEC) of tropical Amazonian soils in the state of Pará, Brazil. The following objectives were set: i) to compare the efficacy of individual and combined Vis-NIR and pXRF data for the prediction of chemical attributes, using the Random Forest (RF) algorithm, and ii) to compare two methods (Boruta and Principal Component Analysis - PCA) for dimensionality reduction. Soil samples were collected from surface and subsurface layers in forest areas. Using data obtained from individual sensors, the pXRF data demonstrated the greatest effectiveness in predicting soil pH for surface (R2 = 0.92; RMSE = 0.27) and subsurface (R2 = 0.92; RMSE = 0.29) samples using Boruta. The most accurate prediction of SOM was achieved using pXRF data with Boruta for surface samples (R2 = 0.92; RMSE = 3.40), while for subsurface samples, the highest accuracy was obtained using Vis-NIR data with Boruta (R2 = 0.85; RMSE = 3.95). The CEC demonstrated the highest degree of accuracy for surface samples using PCA (R2 = 0.89; RMSE = 12.33) and subsurface samples using Boruta (R2 = 0.89; RMSE = 1.96) when utilizing pXRF and Vis-NIR data, respectively. The integration of sensor data yielded more precise predictions for the three variables at both depths using Boruta than with PCA. This study advances the application of proximal sensors for predicting soil fertility in the Amazon and demonstrates their effectiveness.
