Biomass Estimation from Tree Heights on Individual-Level with Gaussian Process Regressor

Abstract

To monitor the forests and estimate the above-ground biomass in national to global scale, remote sensing data have been widely used. However, due to their coarse resolution (hundreds of trees present within one pixel), it’s costly to collect the ground reference data. Thus, an automatic biomass estimation method on individual tree level using high-resolution remote sensing data (such as Lidar data) is of great importance. In this paper, we explored to estimate tree’s biomass from single parameter - the tree height – using Gaussian process regressor. We collected a dataset of 8342 records, in which individual tree’s height (in m), diameter (in cm), and the biomass (in Kg) are measured. Besides, Jucker data with crown diameter measurement are also used. The datasets coverage eight dominant biomes. Using the data, we compared five candidate biomass estimation models, including three single-parameter biomass-height models (proposed Gaussian process regressor, random forest, and linear model in log-log scale) and two two-parameter models (biomass-height-crown diameter model, and biomass height-diameter model). Results showed a high correlation between biomass and height as well as diameter, and the biomass-height-diameter model has low biases of 0.08 and 0.11, and high R-square scores of 0.95 and 0.78 when using the two datasets respectively. The biomass-height-crown diameter has a median performance with R-square score of 0.66, bias of 0.26, and root mean square error of 1.11Mg. Although the biomass-height models are less accurate, the proposed Gaussian regressor has a better performance over linear log-log model and random forest (R-square: 0.66, RMSE: 4.95 Mg; bias: 0.34). Besides, the results also suggest that non-linear models have an advantage over linear model on reducing the uncertainty either when the tree has a large (> 1 Mg) or small (< 10 kg) biomass.