Modeling the productivity of mechanized CTL harvesting with statistical machine learning methods

Modern forest harvesters automatically collect large amounts of standardized work-related data. Statistical machine learning methods enable detailed analyses of large databases from wood harvesting operations. In the present study, gradient boosted machine (GBM), support vector machine (SVM) and ordinary least square (OLS) regression were implemented and compared in predicting the productivity of cut-to-length (CTL) harvesting based on operational monitoring files generated by the harvesters' onboard computers. The data consisted of 1,381 observations from 27 operators and 19 single-grip harvesters. Each tested method detected the mean stem volume as the most significant factor affecting productivity. Depending on the modeling approach, 33-59% of variation was due to the operators. The best GBM model was able to predict the productivity with 90.2% R-2, whereas OLS and the SVM machine reached R-2-values of 89.3% and 87% R-2, respectively. OLS regression still proved to be an effective method for predicting productivity of CTL harvesting with a limited number of observations and variables, but more powerful GBM and SVM show great potential as the amount of data increases along with the development of various big data applications.