Assessing the stability of parameters estimation and prediction accuracy in regression methods for estimating seed oil content in Brassica napus L. using NIR spectroscopy

Brassica napus L., is an oilseed species of great economic importance due to its high oil content in the seed, representing the second worldwide source of edible oil after soybean. To measure the seed oil content a destructive chemical analysis (Soxhlet) is typically used. In addition, Soxhlet is an expensive, time consuming and labor intensive methodology. In order to overcome these drawbacks the use of near infrared spectroscopy (NIR) has been a low cost alternative to determine oil content and other seed quality traits. However, in order to implement accurate NIR based measurements, stable prediction models need to be developed. In the present work, we assess parameters stability using bootstrap and prediction error through Predicted Residual Error Sum of Squares (PRESS) for three methods: Multiple Linear Regression (MLR), Support Vector Regression (SVR) and Artificial Neural Networks (ANN). The results showed that the best behavior of the three methods analized was ANN, where the variance for stability of parameters was 0.027 and the values of PRESS index were 75.65, 226.07 and 314.91 for ANN, MLR and SVR, respectively. These results will contribute to improve the development of regression models for more accurate seed oil content measurements using NIR technology.