Computational analysis of resilient propagation algorithm variants for predicting harvester productivity

Abstract

This study investigated the use of artificial neural networks (ANNs) to estimate the productivity of forestry harvesters operating in commercial eucalyptus plantations in Brazil. The dataset, collected from mechanised harvesting operations over continuous 24-hour shifts in the states of Espírito Santo and Bahia, included quantitative and categorical variables. Quantitative inputs comprised the mean individual tree volume and productivity. Categorical variables included terrain inclination, tree tortuosity and bifurcation, sub-forest presence, operator experience, and type of service supply. Four variants of the Resilient Propagation (Rprop) algorithm were evaluated: Rprop+, iRprop+, Rprop−, and iRprop−. A total of 200 ANN models (50 per variant) were trained using 70% of the data, with the remaining 30% reserved for validation. All networks employed min–max normalisation, sigmoid activation functions, and an eight-neuron hidden layer. Model performance was assessed using the correlation coefficient (R), root mean square error percentage (RMSE%), bias percentage, and variance. The Rprop+ variant achieved the best predictive performance, with R = 0.837, RMSE% = 18.380, bias% = –0.017, and variance = 9.867, proving high reliability in modelling harvester productivity and offering valuable support for planning and decision-making in mechanised forestry operations.