Abstract
Forest biomass estimation at large scale is challenging and generally entails large uncertainty in tropical regions. With their wall-to-wall coverage ability, passive remote sensing signals are frequently used to extrapolate field estimates of forest aboveground biomass (AGB). However, studies often use limited reference data and/or flawed validation schemes and thus report unreliable extrapolation error estimates. Here, we compared the ability of three medium- to high-resolution passive satellite sensors, Landsat-8 (L8), Sentinel-2B (S2) and Worldview-3 (WV3), to map AGB in a forest landscape of Thailand. We used a large airborne LiDAR-derived AGB dataset as a reference to train and validate a random forest algorithm and conducted robust error assessments and variable selection using spatialized cross-validations. Our results indicate that the selected predictors strongly varied among the three sensors and between analyses were restricted to low (≤200 Mg ha−1) and high (>200 Mg ha−1) AGB areas. WV3 and S2 data outperformed L8 data to extrapolate AGB (RMSE of 68 and 72 against 84 Mg ha−1, respectively) due to the inclusion of the red-edge band and, probably, to their higher spatial and spectral resolution. Sensitivity to large AGB values was higher for WV3 than for S2 and L8 with saturation point of 247 Mg ha−1 against 204 and 192 Mg ha−1. AGB values above these saturation points remained poorly predictable, especially for L8, indicating that several tropical forest AGB maps should be interpreted with extreme caution. However, predicted gradients of lower AGB values (≤200 Mg ha−1), i.e., in early forest successional stages, were fairly consistent among sensors (r > 0.70), even if the mean absolute difference between estimates was large when AGB predictions were extrapolated out of the calibration area at regional level (34%). We finally showed that calibrating the model only within the sensitivity AGB domain (e.g., <200 Mg ha−1) minimizes the risk of induced bias for estimating small AGB values. These results provide important benchmarks for interpreting previously published maps and to improve future validation schemes.