Development of a universal NIR model for predicting cellulose and lignin contents in multiple Pinus elliottii clones via efficient feature wavelength selection

Abstract This study developed universal near-infrared (NIR) spectroscopy models for the high-throughput prediction of wood chemical composition across multiple Pinus elliottii clones. To overcome the limited generalizability of conventional full-spectrum models caused by clone-specific spectral variations and redundancy, three feature selection algorithms were evaluated for extracting robust, clone-independent spectral features correlated with cellulose and lignin content. The Monte Carlo successive projections algorithm (MC-SPA) proved most effective, constructing high-performance models using only a minimal subset of the full-spectrum wavelengths. The resulting universal models based on MC-SPA-selected features demonstrated improved predictive accuracy and generalizability across clones. This work highlights the importance of feature wavelength selection in building highly generalizable NIR calibrations and provides a practical strategy for high-throughput screening in multi-clone breeding programs.