Machine learning: Modeling increment in diameter of individual trees on Atlantic Forest fragments

[Display omitted] •Machine learning techniques increase the accuracy of APIdbh predictions.•The P (proxy variable) and SG are the main variables to explain APIdbh variation.•The ANN predicts APIdbh more accurately than SVR and RF. Growth models at individual levels allow a more detailed description of the forest dynamics. However, many fittings using statistical models have shown low performance in native forests due to variation between the increments of different trees or species and even within the same species. In this study, we evaluated the accuracy of predictions of annual periodic increment in diameter (APIdbh; cm yr−1) of individual trees in the Atlantic Forest using three machine learning techniques: artificial neural network (ANN); support vector regression (SVR); and random forest (RF). The study was carried out in three fragments of Atlantic Forest: Ipaba (IP), São José (SJ), and Cachoeira das Pombas (CP) located respectively in the municipalities of Caratinga, Coronel Fabriciano, and Guanhães, Minas Gerais State, Brazil. In IP, SJ and CP were allocated 22, 12 and, 20 plots, respectively, with 0.05 ha each. The data: circumference at 1.30 m (cbh; cm), total height (H; m), and scientific names of trees with cbh > 15 cm were collected in the years 2002, 2007, 2012, and 2017 in the all fragments. Three semi-independent competition indexes were calculated: Glover and Hool (1979), Stage (1973), and adapted from Glover and Hool (1979). In each fragment, the species were grouped into five APIdbh groups through cluster analysis. Seventy percent of the observations of each species, in each period, were set for training and 30% for validation through balanced partitioning. The selection of the best configuration of each technique was through the statistics of correlation between the observed and predicted increments (ryŷ), root mean square error (RMSE), mean absolute error (MAE), weighted value (WV), observed and predicted APIdbh scatter plots and scatter plots of the standardized residuals as a function of diameter at the height of 1.30 m (dbh). After selecting the best configuration of each technique in each fragment, cross-validation was applied with 10 partitions and 50 repetitions. The RMSE means were calculated for each of the 50 cross-validation repetitions. Friedman and Nemenyi nonparametric tests were applied to evaluate the best technique for each fragment, based on RMSE means in cross-validation. The ANN presented the lowest RMSE mean (0.1382