Comparing spectral models to predict manganese content in Rosa spp. leaves using VIS-NIR data

This study, conducted on the Freedom rose cultivar grown under greenhouse conditions in the municipality of Tocancipá, Cundinamarca (Colombia), implemented Partial Least Squares Regression (PLSR) and Principal Component Regression (PCR) methods using visible and near infrared (VIS-NIR) spectroradiometry from 350 to 2500 nm to predict manganese (Mn) content in rose leaves. A randomized complete block design (RCBD) with manganese doses of 0%, 25%, 50%, 75%, and 100% of the reference dose of 2 mg L-1 was established in 25 plots with five treatments and five replicates. Samplings were conducted in the five phenological stages of “palmiche”, “rice”, “chickpea”, “scratch color”, and “straight sepals”, analyzing 10 plants per treatment, and spectral responses were measured on the adaxial leaf surface using the FieldSpec® 4 spectroradiometer. For model generation (PLSR and PCR), 24 predictive models were evaluated, comprising three spectral response ranges: range 1 (350-1000 nm), range 2 (350-1800 nm), and range 3 (350-2500 nm), applying different spectral correction methods: raw data (RD), Savitzky-Golay (SG), range normalization (RN), and Savitzky-Golay followed by range normalization (SG-RN). A total of 100 samples were used: 80 for calibration and 20 for external validation, randomly selected to represent the variability of the treatments. The spectral corrections improved the accuracy and robustness of the predictions, with the RN-PLSR and SG-RN-PLSR models showing the best performance metrics (R², RMSE, and RPD). The most relevant wavelengths were 523 nm, 557 nm, and around 720 nm, with correlations greater than 0.6 with the Mn concentration in leaves.