Data Fusion of Visible-Near-Infrared and Mid-Infrared Spectra for Predicting Key Soil Properties across Different Soil Layers

Abstract

Nutrient availability for crops is not limited to the topsoil; deeper layers also serve as significant sources. Rapid determination of soil properties at different depths is essential for the precision management of farming inputs. Most existing studies focus on predicting surface soil properties using spectral data from a single sensor, while research exploring the potential of multisensor spectral fusion for soil analysis at varying depths remains limited. This study evaluated the predictive performance of: (1) single-spectrum modeling using visible and near-infrared (vis–NIR) and mid-infrared (MIR) spectrophotometers and (2) two spectral fusion methods─direct concatenation and outer product analysis of full-spectral absorbance─for five key soil properties: pH, total organic carbon (TOC), available phosphorus (AP), available potassium (AK), and magnesium (Mg), across three soil depths. The data set comprised 176 fresh soil samples collected from 59 locations across five arable cropping fields. Prediction models were developed using partial least-squares regression (PLSR) and support vector machine (SVM), whose performance was assessed using the coefficient of determination (R²), root-mean-square error (RMSE), and ratio of performance to interquartile distance (RPIQ). Results showed that vis–NIR spectra generally outperformed MIR, with validation R² ranging from 0.39 to 0.67 and RPIQ from 0.84 to 3.25, compared to MIR (R² = 0.42–0.62, RPIQ = 0.97–3.08). Notably, spectral fusion using the OPA–SVM method yielded the best predictions for TOC (R² = 0.75, RPIQ = 3.35) and AP (R² = 0.83, RPIQ = 4.72), while the DC-PLSR model achieved the highest performance for pH (R² = 0.65, RPIQ = 2.43). However, fusion was not always superior to single-spectrum models; for example, vis–NIR-SVM and MIR-SVM gave the best results for AK (R² = 0.53, RPIQ = 1.45) and Mg (R² = 0.61, RPIQ = 1.30), respectively. Given these varying results, we recommend selecting spectroscopic techniques based on both predictive performance and practical considerations such as cost-effectiveness and operational feasibility.