Accepted: 2025-09-29
The timely and accurate assessment of soil nutrient information is essential for ensuring global food security and sustainable agricultural development. This study evaluated the individual and fusion performance of mid-infrared (MIR) and portable X-ray fluorescence (pXRF) spectroscopy for predicting selected soil properties. Four sensor fusion strategies were implemented: direct concatenation (DC), feature-level fusion using stability competitive adaptive reweighted sampling (sCARS) and least absolute shrinkage and selection operator (LASSO) algorithms (sCARS-C and LASSO-C), multi-block fusion via sequential orthogonal partial least squares (SO-PLS), and Granger-Ramanathan model averaging (GRA) method to enhance prediction accuracy for 13 soil properties. The findings revealed that single sensor models using either MIR or pXRF provided accurate estimations for soil organic matter (SOM), total nitrogen (TN), available phosphorus (AP), calcium (Ca), iron (Fe), manganese (Mn), and pH, but showed limitations for total potassium (TK), magnesium (Mg), copper (Cu), zinc (Zn), available potassium (AK), and total phosphorus (TP). The DC model significantly improved predictions for Mg (Rp2=0.76, RMSEp=358.76 mg kg-1, RPDp=2.03) and TK (Rp2=0.75, RMSEp=775.96 mg kg-1, RPDp=2.00). The LASSO-C model demonstrated superior prediction accuracy compared to the DC model for AP, AK, TP, Zn, Mn, and Cu, achieving optimal results for AP (Rp2=0.89, RMSEp=21.37 mg kg-1, RPDp=3.01) and Zn (Rp2=0.80, RMSEp=9.88 mg kg-1, RPDp=2.32). This enhancement is attributed to LASSO's effective selection of feature information from the complete MIR and pXRF spectra. The GRA models achieved the highest prediction accuracy for TP, pH, AK, and Cu, with Rp2 values of 0.80, 0.82, 0.82, and 0.65, RMSEp values of 129.21 mg kg-1, 0.13, 48.38 mg kg-1, and 3.87 mg kg-1, and RPDp values of 2.23, 2.34, 2.37, and 1.67, respectively. For single-sensor applications, MIR spectra are recommended for predicting SOM, TN, and Ca (Rp2≥0.88, RPDp≥2.87), while pXRF is more cost-effective for measuring Ca, Fe, and Mn (Rp2≥0.80, RPDp≥2.22). This research demonstrates the effectiveness of MIR and pXRF sensor fusion in enhancing soil nutrient assessment accuracy, particularly for available nutrients and micronutrients.