Utilization of neural networks for the estimation of aboveground forest biomass from Ikonos satellite image and multi-source geo-scientific data

The present study develops an approach for the estimation of aboveground forest biomass based on neural networks, using Ikonos satellite image data and multi-source geo-scientific data. Two methods of aboveground forest biomass estimation were compared: multiple regressions and the neural networks. Percentages of residual errors of the neural networks biomass estimates were lower than 1% for all groups of species, except for “intolerant hardwood” which had a percentage of 3.2% for the 9-18 cm DBH class. Percentages of residual errors of biomass estimates were higher with the quadratic multiple regression approach than with the neural networks, particularly for “intolerant hardwood” where a value of 51.41% was observed for the 19–40+ cm DBH class. Root mean square error values (RMSE) calculated from biomass estimates resulting from the neural networks approach were lower than those computed with estimates of the quadratic multiple regressions model, for all groups of species.