Predicting the strength of Populus spp. clones using artificial neural networks and ε-regression support vector machines (ε-rSVM)

Abstract

Wood properties, including bending stiffness and strength, basic density and microfibril angle were experimentally obtained for six aspen and six hybrid poplar clones grown in Western Canada. Data analysis attempted to establish a relationship between wood mechanical properties and intrinsic wood attributes by means of artificial neural networks (ANN) and ε-regression support vector machines (ε-rSVM) employing a 5-fold cross validation approach (5-fold CV). Initial results for strength were acceptable, but require further improvement. Estimations of stiffness results (MOE) were inferior to those of strength (MOR) due to the fact that in several regression cases, the developed model worked well for narrow windows of data, but failed on a large scale due to the high variations in the values of the input data vectors. In such cases, the result is probably the development of regression with uneven performance throughout the input data set, and therefore the modeling capacity is poor. To avoid this predicament, different neural networks with one output neuron were developed in order to estimate either the stiffness or the strength, and at the same time the approximation capabilities of ε-rSVM were employed. In both methods, 5-fold CV was carried out in order to attain a more generalized solution by eliminating the boundary effect phenomena and by avoiding local behavior of the global support vector regression. The resultant models were evaluated by common metrics. The best ANN for the estimation of strength in combination with 5-fold CV, was a modular back propagation with average R²=0.70, and mean root mean square error (MRMSE) equal to 0.19 and mean average percent error (MAPE) equal to 12.5%. The Gaussian kernel 5-fold CV ε-rSVM estimated MOR with similar accuracy. The best 5-fold CV ANN for MOE estimation was a feed forward back propagation one, with average R²=0.60, MRMSE equal to 0.23 and MAPE equal to 41.5%, which was better than all other kernel methods employed.