ABSTRACT Productivity analysis in mechanized harvesting has traditionally relied on statistical expertise and mathematical modeling. However, machine learning tools have emerged as a viable alternative, as they serve the same purpose, utilizing a combination of varied attributes (quantitative and qualitative) and handling large datasets. This study aimed to determine whether the inherent attributes of mechanized timber harvesting of Eucalyptus spp. plantations enable the creation of a high-performance model that can accurately predict productivity from machine learning. For the modeling, we considered five attributes concerning forest inventory, in addition to working hours and the operator experience level. We considered the productivity, timber harvested per working hour, as the target attribute of the modeling. We subjected the database to 17 common algorithms in default mode and compared them according to error metrics and accuracy. We also determined the relative importance of each attribute in the predictive model. The inherent attributes concerning mechanized timber harvesting of Eucalyptus spp. plantations evaluated in this study enable the creation of a high-performance model that can accurately predict productivity from machine learning. The Gradient boosting model in ensemble mode can predict the productivity of harvesters in Eucalyptus spp. plantations with an R2 of 0.81. The attributes that have greater relative importance are operator experience level, average individual tree volume, and stand density with 100%, 76.3%, and 65.8%, respectively.
Read full abstract