Bee-inspired Radial Basis Function network to estimate tree aboveground wood volume in Brazilian Savanna

Forest variables such as aboveground wood volume are usually estimated by applying regression analysis. Nevertheless, in the past decades, new alternatives have been used, for example, artificial neural networks. In this sense, our study aimed to evaluate the efficiency of the Radial Basis Function (RBF) neural network in estimating aboveground wood volume for the Brazilian savanna using the cOptBees training algorithm. We fitted 18 allometric models and trained three Multilayer Perceptron (MLP) networks and two RBF networks using two different algorithms: k-means and cOptBees. We selected the MLP and RBF networks, as well as the allometric model with the best accuracy, for comparison. We verified the methods’ accuracy by analysing the statistics of bias, root mean square error (RMSE), and Pearson’s correlation coefficient. The lowest bias value was presented by the RBF network using the cOptBees algorithm (5.90 x 10-5). The Näslund model showed the highest correlation (9.45 x 10-1), as well as the lowest RMSE (2.07 x 10-3). The aboveground wood volume estimates provided by the artificial neural networks showed similar results to those provided by classical regression models. Overall, we may infer that RBF networks trained using the cOptBees algorithm can be used to estimate aboveground wood volume in the Brazilian savanna, being as accurate as MLP or RBF networks trained with the k-means algorithm and allometric models. Finally, all methods showed similar aboveground wood volume estimates. However, neural networks offer advantages in optimizing field surveys because they require less sampling effort.