Improving Tropical Forest Biomass Predictions with Multi-Output Deep Learning Models for Above- and Belowground Estimates

Abstract

Accurate biomass prediction in tropical forests is a key factor in understanding carbon sequestration and forest health. However, traditional methods lack the capability to deal with complex ecological data. The current study seeks to evaluate the Extra Trees algorithm as a novel optimization technique to improve the reliability of simultaneous predictions for aboveground biomass, belowground biomass, and total biomass in tropical forests, and to perform against other machine learning models derived from a dataset of 175 tree samples collected across 27 plots in the Central Highlands ecoregion, Vietnam. It also compares the performance of ET against other machine learning models such as GB, Bagging, and LSTM networks. Our findings showed that ET had the best performance regarding prediction accuracy, since the training VAFs obtained were 99.90391 for BGB, 99.87924 for AGB, and 99.88903 for TB when combined with R-squared (R²) values of 0.999039, 0.998792, and 0.99889, respectively. Additionally, ET was found to have lower MAE and RMSE values than the results from other models tested in this work. This makes ET a very useful tool for biomass estimates. This result acknowledges the potential of ET for giving more accurate and reliable biomass prediction than traditional methods; thus, it would be contributing significantly to forest management and conservation strategies.