Application of Machine Learning for Aboveground Biomass Modeling in Tropical and Temperate Forests from Airborne Hyperspectral Imagery

Abstract

Accurate operational methods used to measure, verify, and report changes in biomass at large spatial scales are required to support conservation initiatives. In this study, we demonstrate that machine learning can be used to model aboveground biomass (AGB) in both tropical and temperate forest ecosystems when provided with a sufficiently large training dataset. Using wavelet-transformed airborne hyperspectral imagery, we trained a shallow neural network (SNN) to model AGB. An existing global AGB map developed as part of the European Space Agency’s DUE GlobBiomass project served as the training data for all study sites. At the temperate site, we also trained the model on airborne-LiDAR-derived AGB. In comparison, for all study sites, we also trained a separate deep convolutional neural network (3D-CNN) with the hyperspectral imagery. Our results show that extracting both spatial and spectral features with the 3D-CNN produced the lowest RMSE across all study sites. For example, at the tropical forest site the Tortuguero conservation area, with the 3D-CNN, an RMSE of 21.12 Mg/ha (R² of 0.94) was reached in comparison to the SNN model, which had an RMSE of 43.47 Mg/ha (R² 0.72), accounting for a ~50% reduction in prediction uncertainty. The 3D-CNN models developed for the other tropical and temperate sites produced similar results, with a range in RMSE of 13.5 Mg/ha–31.18 Mg/ha. In the future, as sufficiently large field-based datasets become available (e.g., the national forest inventory), a 3D-CNN approach could help to reduce the uncertainty between hyperspectral reflectance and forest biomass estimates across tropical and temperate bioclimatic domains.