Global scale-biomass estimation based on a deep learning method

Abstract

Modeling accurate above-ground biomass (AGB) maps is a critical issue in remote sensing research. Since the relationship between biomass and environmental variables are usually complex, because of being affected by many factors, using non-parametric methods like Convolutional Neural Network (CNN) to estimate biomass on the global scale is convenient. To choose the most significant variables to enter to the AGB estimation model two feature selection techniques were applied, Support Vector Machine for Regression Feature Selection (SVRFS) and Random Forest Feature Selection (RFFS) techniques. The optimum AGB model was created using the training dataset and the predicted model was created using the test dataset. The results showed CNN with the SVRFS technique, achieved the highest RMSE values (31.22 Mg/ha). This study highlighted the capability of the deep learning algorithm to improve AGB estimates on a global scale.