Unveiling the potential of Brachiaria ruziziensis: Comparative analysis of multivariate and machine learning models for biomass and NPK prediction using Vis-NIR-SWIR spectroscopy

Abstract

This study investigated the development and validation of predictive models for estimating foliar nitrogen (N), phosphorus (P), and potassium (K) contents, along with shoot dry mass (SDM) of Brachiaria ruziziensis L. The approach utilized Vis-NIR-SWIR spectroscopy coupled with multivariate statistical techniques (PLS, PCR) and machine learning algorithms (SVM, RF). A triple-factorial, completely randomized design with ten replications per treatment was employed in a greenhouse setting. Treatments included type of input (limestone-mining coproducts), input particle size (filler and powder), and soil class (Arenosol and Ferralsol). Following input incubation, B. ruziziensis was sown. Forty days later, foliar spectra and leaves were collected. Chemical analysis determined NPK content, along with SDM. The study developed predictive models utilizing Vis-NIR-SWIR spectroscopy, Partial Least Squares (PLS), and machine learning algorithms like Support Vector Machine (SVM) and Random Forest (RF) to estimate foliar N, P, K, and biomass. Model adjustments achieved R²_p > 0.70 and RPD_p > 1.80 for PLS, SVM, and RF models across all variables (SDM, N, P, and K). These results highlight the effectiveness of specific spectral bands for nutrient and biomass discrimination and emphasize the potential of these techniques for rapid, non-destructive nutrient content estimation. The findings support the integration of advanced spectroscopic methods with machine learning algorithms for improved precision agriculture practices, providing a more sustainable alternative for nutrient and biomass analysis in forage crops. This approach optimizes forage production and minimizes atmospheric CO₂ emissions.