Bio-inspired algorithm for variable selection in i-PLSR to determine physical properties, thorium and rare earth elements in soils from Brazilian semiarid region

Studies regarding physical parameters and Rare Earth Elements (REEs) in soils are essential for agriculture and raw material extraction. The determination of such parameters is costly and time-consuming. On the other hand, Near-Infrared Spectroscopy (NIRS) is a low cost, non-destructive, and fast alternative to predict REEs and physical properties in soils. Moreover, the specific wavelengths or intervals correlated to target properties are not clearly identified in the spectra, drawing attention to using the variable selection algorithm to resolve this problem. In this sense, this study aimed to develop and test the bio-inspired variable selection method known as Firefly intervals Partial Least Square Regression (FF-iPLSR). This heuristic approach is based on the attractiveness between fireflies as swarm behavior when searching for food. The performance of the FF-i-PLSR was evaluated using six soil profiles from the Brazilian semiarid region using raw data and the preprocessing MSC, SNV, baseline adjusts, and Savitzky-Golay derivative. The following variables were quantified: sand, silt, clay, Th, and total REE. Comparatively, four chemometrics methods were used: full-spectrum PLS, PLS based on variable intervals (iPLS), and the Successive Projections Algorithm for iPLS (iSPA-PLS). The proposed algorithm for variable interval selection obtained a well-adjusted and more parsimonious chemometric model, mainly for silt, sand, and clay, with lower LV's and variable intervals than the comparative models. These results indicated the robustness of the proposed algorithm. The models' quality parameters showed that FF-iPLS achieved low average values for RMSEP, bias, and REP. Thus, the proposed algorithm of variables interval selection FF-iPLS is an interesting tool for analyzing environmental matrices such as soil. However, more studies are necessary to evaluate the algorithm's efficiency using other samples and environmental conditions.