Towards Optimal Variable Selection Methods for Soil Property Prediction Using a Regional Soil Vis-NIR Spectral Library

Soil visible and near-infrared (Vis-NIR, 350–2500 nm) spectroscopy has been proven as an alternative to conventional laboratory analysis due its advantages being rapid, cost-effective, non-destructive environmentally friendly. Different variable selection methods have used deal with the high redundancy, heavy computation, model complexity of using full spectra in spectral modelling. However, most previous studies a linear algorithm selection, application non-linear remains poorly explored. To address current knowledge gap, based on regional soil Vis-NIR library (1430 samples), we evaluated seven algorithms together three predictive predicting properties. Our results showed that Cubist overperformed partial least squares regression (PLSR) random forests (RF) properties (R2 > 0.75 for organic matter, total nitrogen pH) when spectra. Most can greatly reduce number bands therefore simplified models without losing accuracy. The also there was no silver bullet optimal among different algorithms: (1) competitive adaptive reweighted sampling (CARS) always performed best PLSR algorithm, followed by forward recursive feature (FRFS); (2) elimination (RFE) genetic (GA) generally had better accuracy than others algorithm; (3) FRFS performance RF algorithm. In addition, matched outcome this study provides valuable reference information spectroscopic techniques algorithms.