Prediction of lignin content in sugarcane from fresh leaf using near infrared spectroscopy and chemometric methods

INTRODUCTION: Nowadays there is a global concern for reducing the fossil fuels consumption like oil, coal and natural gas. In this context, Brazil has a prominent role on the international scene due to concerns about energy security. Studies have shown that the sugarcane bagasse is chemically feasible for the production of chemical fuels and materials. Among the components of lignocellulosic waste, lignin is strategically important; either for energy cogeneration by burning or production of chemical inputs. Thus, the determination of lignin is an important variable in genetic breeding. However, the lignin determination is a time consuming procedure, expensive and not environmentally friendly. Approaches based on near infrared spectroscopy (NIR) are simple, fast, present relative low cost, wide application and are environmentally friendly.
Lignin prediction in sugarcane bagasse from leaf NIR spectrum using partial least squares (PLS) and features selection is the goal of this work.
EXPERIMENTAL: Lignin content was determined in 256 genotypes, using dry and milled bagasse and applying the Klason method. Lignin values were in a range from 18.05 to 28.37%. The independent variables (NIR spectra) were obtained from Varian FTIR 660 in the range of 10000 - 4000 cm-1, direct from the +3 leaf without any sample preparation. Sugarcane leaves are alternate and are attached to the stalk, with one leaf per internode. Therefore, it was used the middle third of the sheet, excluding the midrib. The proposed work dedicated to correlate the leaf spectrum with the bagasse lignin content, considering the same genotype.
The Kennard Stone algorithm was used to select 216 and 40 samples for calibration and prediction sets, respectively. Furthermore, the models were built using the partial least squares regression (PLS). Different algorithms for features selection were tested i.e., interval PLS (iPLS), genetic algorithm (GA) and the ordered predictors selection method (OPS). All calculation were performed using home-built functions written for Matlab.
RESULTS AND DISCUSSION: Several preprocessing methods were applied but the second derivative showed better performance. The quality of the models is assessed by the root mean square error (RMSE), the correlation coefficient R and ratio of performance deviation (RPD). When internal validation (cross validation—CV) is applied, the error and correlation coefficients are named RMSECV and Rcv, respectively. For external validation (predicting samples – P) the error and correlation coefficients are named RMSEP and Rp, respectively. RMSECV, RPD, Rcv, RMSEP and Rp values obtained for the model were, respectively: Full (1.32, 1.41, 0.78, 0.77, 0.93); GA (0.72, 2.72, 0.94, 0.64, 0.95); iPLS (1.84, 1.15, 0.62, 0.99, 0.90) and OPS (0.76, 2.56, 0.93, 0.67, 0.96). Relative error in prediction for PLS-GA or PLS-OPS models were around 2.5%. Although GA has presented slightly better results, the OPS algorithm is considerably simpler and faster. In addition, the OPS algorithm selected a lower number of variables with a higher predictive capacity. Results confirm that the PLS-OPS model showed high accuracy to predict the lignin content in sugarcane, reducing significantly the time spent with the wet analysis, the cost and the chemical reagents consumption.