Evaluating the Marginal Contribution of Remote Sensing for Forest Biomass Estimation When Inventory Data Exists

Combining remote sensing data with field inventory data is a common practice in estimation of forest Aboveground Biomass (AGB). However, in areas with well-established ground monitoring systems, the actual added value of this combination has not been fully quantified. This paper aims to critically assess the marginal contribution of optical remote sensing data to AGB estimation when the forest inventory data are already available, and further investigates error sources and residual distributions. Using Longyan City, Fujian Province as a case study, we systematically tested the effects of data types (inventory only, Landsat 8 only, inventory + Landsat 8, and inventory + Landsat 8 + meteorological data), sampling methods, and statistical models on plot-scale AGB estimation accuracy, along with uncertainty distribution across biomass levels. Inventory data alone (stand age and canopy cover) achieved acceptable accuracy, with R2 = 0.60 and RMSE = 35.78 t/ha under the optimal configuration (RandomForest + ShuffleSplit_5). Adding Landsat 8 data yielded only modest improvements: RMSE decreased by 5.3% and R2 increased by 6.7% relative to the inventory-only baseline. When the optimal combination for each data type was evaluated on the independent test set, the highest R2 reached only 0.56, leaving approximately 44% of the observed variation in plot-level AGB unexplained. ANOVA showed that data type was the dominant factor, accounting for 73.0% of the variation in RMSE, followed by model choice (15.1%) and their interaction (8.5%), whereas the contribution of the validation scheme was less than 3.0%. We further examined residual distributions across biomass levels and found that residual skewness shifted systematically: negative skewness (overestimation) prevailed in low-biomass plots, near-zero skewness in medium-biomass plots, and positive skewness (underestimation) in high-biomass plots. Among all data types, the Landsat-only configuration exhibited the most severe bias at both low and high biomass extremes; adding remote sensing and climate data to inventory data did not substantially correct this bias structure. Overall, these findings demonstrate that in regions with high-quality ground inventory, the marginal gains from integrating optical remote sensing are limited—both in terms of overall accuracy and the structure of prediction errors. The unexplained variance and systematic residual biases across biomass gradients highlight the need for better representation of high-biomass stands and suggest that future efforts should prioritize sample augmentation in under-represented biomass classes, rather than relying solely on multisource data fusion for accuracy improvement.