Data fusion of XRF and vis-NIR using p-ComDim to predict some fertility attributes in tropical soils derived from basalt

Abstract

X-ray fluorescence (XRF) and visible and near-infrared (vis-NIR) spectroscopies have been capable of predicting soil properties. Both techniques enable environmentally friendly soil analysis with minimal or no sample preparation, providing information about inorganic and organic-mineral soil components. Recent studies have demonstrated that XRF and vis-NIR data fusion may improve the quality of predictive models for soil attributes. Although promising results have been related to the vis-NIR and XRF data fusion for soil analysis, no approach has been consistently optimal. In some cases, the data fusion worsened the models' performance compared with individual models. Moreover, most of these studies are focused on model performance. Few studies assess the synergism among different techniques and how their complementarities/differences may improve the knowledge about the sample spectral responses. For this purpose, this study aimed to assess p-ComDim analysis as a mid-level data fusion approach for soil fertility prediction using vis-NIR and XRF data. Complementarily, p-ComDim results were compared with vis-NIR and XRF individual models and low- and high-level data fusion approaches. Soil organic carbon (SOC), exchangeable calcium (Ca2+) and magnesium (Mg2+), pH, the sum of base (SB), and cation exchange capacity (CEC) were considered representative variables of the soil chemical fertility. A relative improvement ranging from 5% to 33% was observed comparing p-ComDim with vis-NIR and XRF individual models. Moreover, p-ComDim results were better when compared with the two data fusion approaches most used in literature for soil fertility prediction (low- and high-level data fusion). In addition, the p-ComDim analysis allowed extracting the weights (saliences) of each technique as well as the synergism among their variables to predict each attribute, demonstrating that the most important common dimensions (CDs) to the predictive models in all analyzed attributes were those with higher silences in the XRF block. It shows the synergism among different spectral ranges and the potential of p-ComDim for vis-NIR and XRF data fusion in soil fertility prediction. Therefore, the results from the present study may contribute to a better understanding between soil sample constituents and spectral responses, assisting in the explainability of the multivariate regression models.