Comparison of multi-source remote sensing data for estimating and mapping above-ground biomass in the West Usambara tropical montane forests

Abstract

Above-ground biomass (AGB) estimation is important to better understand the carbon cycle and improve the efficiency of forest policy and management activities. AGB estimation models, using a combination of field data and remote sensing data, can largely replace traditional survey methods for measuring AGB. There are, however, critical steps for mapping AGB based on satellite data with an acceptable degree of accuracy, such as choice of remote sensing data, the proper statistical modelling method, and remote sensing predictor variables, at known field locations. This study sought to identify the optimal optical and synthetic aperture radar (SAR) remote sensing imagery from five sensors (PlanetScope, Sentinel-2, Landsat 8 OLI, ALOS-2/PALSAR-2, and Sentinel-1) to model 159 field-based AGB values from two montane forests under semiparametric (Generalized Additive Model; GAM) and non-parametric (eXtreme Gradient Boosting; XGB) approaches using information from four groups of predictor variables (spectral bands/polarizations, vegetation indices, textures, and a combination of all). The study's results showed that PlanetScope (rRMSE = 69.19%; R2 = 0.161) was the most precise optical sensor while ALOS-2/PALSAR-2 (rRMSE = 70.76; R2 = 0.165) was the most precise amongst the SAR sensors. XGB models generally resulted in those with lower prediction errors as compared to GAMs for the five sensors. Models having textures of vegetation indices and polarization bands achieved greater accuracy than models that incorporated spectral bands/polarizations and vegetation indices only. The study recommends that PlanetScope and ALOS-2/PALSAR-2 remote sensing data using the XGB-based technique is an appropriate approach for accurate local and regional estimation of tropical forest AGB particularly for complex montane forest ecosystems.