Evaluation of machine learning methods and multi-source remote sensing data combinations to construct forest above-ground biomass models

Abstract

Rapid and accurate estimation of forest biomass are essential to drive sustainable management of forests. Field-based measurements of forest above-ground biomass (AGB) can be costly and difficult to conduct. Multi-source remote sensing data offers the potential to improve the accuracy of modelled AGB predictions. Here, four machine learning methods: Random Forest (RF), Gradient Boosting Decision Tree (GBDT), Classification and Regression Trees (CART), and Minimum Distance (MD) were used to construct forest AGB models of Taiyue Mountain forest, Shanxi Province, China using single and multi-sourced remote sensing data and the Google Earth Engine platform. Results showed that the machine learning method that most accurately predicted AGB were GBDT and spectral index for coniferous (R2 = 0.99; RMSE = 65.52 Mg/ha), broadleaved (R2 = 0.97; RMSE = 29.14 Mg/ha), and mixed-species (R2 = 0.97; RMSE = 81.12 Mg/ha) forest types. Models constructed using bivariate variable combinations that included the spectral index improved the AGB estimation accuracy of mixed-species (R2 = 0.99; RMSE = 59.52 Mg/ha) forest types and reduced slightly the accuracy of coniferous (R2 = 0.99; RMSE = 101.46 Mg/ha) and broadleaved (R2 = 0.97; RMSE = 37.59 Mg/ha) forest AGB estimation. Overall, parameterizing machine learning algorithms with multi-source remote sensing variables can improve the prediction accuracy of mixed-species forests.