The effect of class-balance and class-overlap in the training set for multivariate and product-adapted grading of Scots pine sawn timber

Using multivariate partial least squares regression (PLS) to perform visual quality grading of sawn timber requires a training set with known quality grades for the training of a grading model. This study evaluated the grading accuracy of an independent test set of sawn timber when changing the aspects of class-balance and class-overlap of the training set consisting of 251 planks. The study also compared two ways of expressing the reference-grade of the training set; by grading images picturing the planks, and by grading the product produced from the planks. Two grading models were trained using each reference-grade to establish a baseline for comparison. Both models achieved a 76% grading accuracy of the test set, indicating that both reference-grades can be used to train comparable models. To study the class-balance and class-overlap aspects of the training set, 25% of the training set was removed in two training scenarios. The models trained on class-balanced data indicated that class-imbalance of the training set was not a problem. The models trained on data with less class-overlap using the product-grade reference suffered a 4%-points grading accuracy loss due to the smaller training set, while the model trained using the image-grade reference retained its grading accuracy.