BIOMASS ESTIMATION BASED ON MULTILINEAR REGRESSION AND MACHINE LEARNING ALGORITHMS IN THE MAYOMBE TROPICAL FOREST, IN THE DEMOCRATIC REPUBLIC OF CONGO

Abstract

Accurate forest aboveground biomass estimations have always been of crucial importance for sustainable forest management. However, a choice of the suitable statistical modelling method and predictor variables from remotely sensed data remains the keystone for providing accurate aboveground biomass estimates. The present study intended to compare the potential of four modelling techniques, including RandomForest, Support vector machine, multilinear regression, and K-nearest neighbour for estimating aboveground biomass using vegetation indices, spectral information, and both vegetation indices and spectral bands. The results have revealed that machine learning algorithms provide better results than the multilinear regression method. Indeed, the multilinear regression method produced the lowest R 2 and the greatest RMSE. Besides, the RandomForest performed better by providing accurate results compared to other machine learning algorithms. However, comparing the three sets of predictors, the vegetation indices have yielded accurate results of aboveground biomass and the strongest modelling power. Our results have also revealed that the RF is the best choice for predicting aboveground biomass for the purpose of reducing over-or under-estimation problems. This study has demonstrated the potential of the machine learning algorithms in predicting aboveground biomass in the tropical forest, using freely remotely sensed data derived from sensors with medium spatial resolution.