Enhancing soil property predictions using spectral fusion: Comparisons between outer product analysis and vector concatenation and among modeling algorithms

Abstract

Multiple approaches have been employed to improve predictions of soil properties using vis-NIR-SWIR and mid-IR spectra, including techniques for extracting/enhancing spectral features and various types of regression models. However, only a few studies have conducted statistical comparative analyses to evaluate the impact of spectral fusions and different algorithms in enhancing prediction performances. So, the aim of this study was to compare the predictions of several soil properties (32 in total) using individual (vis-NIR-SWIR and mid-IR) and fused spectra and four different modeling algorithms. We used a spectral database comprising 1259 soil samples, collected in four Brazilian states, and containing chemical, physical, and mineralogical attributes. Outer Product Analysis (OPA) and Vector Concatenation, called here Side-by-Side (SbS), were used as spectral fusion strategies. Additionally, we also tested the following regression models: Support Vector Machine (with linear and radial Kernel functions), Partial Least Squares Regression, and Boosted Regression Trees. Our results showed the effectiveness of both fusion strategies in enhancing the spectroscopic modeling of soil properties compared to the individual ranges. Fused spectra significantly improved the modeling of most soil properties by utilizing all spectral information in distinct ways. Comparing these two strategies, OPA produced better results (models with R² ≥ 0.50) for 70 % of all soil properties, while this proportion was 48 % for SbS. Support Vector Machine with linear kernel trick performed the best significant modeling results for 78 % of all soil properties with R² ≥ 0.50. Thereby, our results confirm the usefulness of the spectral fusion as an effective technique to improve the spectroscopic prediction of soil properties important for soil survey, classification, mapping, and management.