XRF and gamma-ray data fusion for predicting key soil fertility attributes

Abstract

Efforts to address the limitations of traditional wet chemistry methods have driven studies combining proximal soil sensors with machine learning (ML) to assess soil health through fertility attributes. While energy dispersive X-ray fluorescence (XRF) has been widely and successfully applied, gamma-ray spectrometry (GR) remains relatively underexplored. At the same time, XRF and GR integration (XRF + GR) by data fusion techniques represent a promising pathway to increase the reliability of predictive results. From this perspective, this study aimed to address scientific gaps on deepening data fusion approaches, applying more complex ML algorithms beyond traditional linear regression methods, and expanding findings to encompass a broader set of fertility attributes. Specifically, we compared the standalone performance of XRF vs. XRF + GR by low- and high-level data fusion strategies employing Random Forest (RF), Partial Least Squares (PLS), and Artificial Neural Networks (ANN-MLP) models for assessing soil organic carbon (SOC), cation exchange capacity (CEC), plant-available Magnesium (avMg), Potassium (avK), and Calcium (avCa). A Rhodic Ferrasol soil dataset (n=80) was exploited. Although RF achieved the highest predictive performance compared to PLS and ANN-MLP (0.45 ≤ R² ≤ 0.78, 1.29 ≤ RPIQ ≤2.68) on XRF data, they were outperformed by data fusion approaches for most attributes, as the relative improvement indicated enhancements for SOC (12.3 %), CEC (20.2 %), and avCa (20.9 %) modeling. While low-level fusion models attained excellent results for SOC (R² = 0.83, RPIQ = 3.05) and avCa (R² = 0.82, RPIQ = 3.19), high-level fusion excelled in CEC prediction (R² = 0.70, and RPIQ = 2.83). The standalone XRF approach proved sufficient for avK modeling (R² = 0.71, and RPIQ = 2.52), however, avMg prediction remained challenging across all methods. Overall, the XRF + GR synergy enhanced the predictive accuracy of most evaluated soil parameters (SOC, CEC, and avCa), providing valuable insights for developing more efficient, precise, and sustainable soil monitoring methodologies at local scale.