Comparative Multi-Approach Mapping of GEDI-Derived Forest Biomass Classes Using Unsupervised Learning, Supervised Modelling and AHP-Based Multicriteria Analysis in the Luki Biosphere Reserve, DRC

Introduction: Wall-to-wall estimation of tropical forest biomass remains challenging because GEDI provides accurate LiDAR-derived above-ground biomass density (AGBD) observations only at discrete footprints, while optical and radar predictors may show limited sensitivity in dense forests. This study evaluates the capacity of multi-source Earth observation predictors to reproduce GEDI-derived biomass patterns in the Luki Biosphere Reserve, Democratic Republic of the Congo.
Methods: A total of 2,160 valid GEDI AGBD observations were analysed using predictors derived from Sentinel-1 radar, Sentinel-2 optical bands, vegetation and moisture indices, texture metrics, Dynamic World land-cover probabilities and DEM-based topographic variables. The workflow combined an initial 27-predictor exploratory phase with principal component analysis and unsupervised classification, an Enhanced dataset containing 180 usable predictors, supervised modelling, predictor-discrimination analysis and AHP-based multicriteria classification using 12 selected criteria. To assess whether target simplification improved separability, both four-class mapping and binary Low/High biomass stratification were tested. Continuous AGBD regression was also evaluated. The revised modelling framework included 21 binary classification trials and 21 regression trials, covering 3 model families, 7 predictor sets and spatially independent 5-fold cross-validation.
Results: Four-class classification with all 180 predictors achieved 0.346 overall accuracy, 0.341 macro-F1 and 0.128 Cohen’s kappa. The AHP multicriteria classification reached 0.335 overall accuracy, 0.338 macro-F1, 0.114 kappa and 0.738 within-one-class agreement. Binary biomass stratification improved discrimination, reaching 0.609 accuracy, 0.609 balanced accuracy, 0.609 macro-F1, 0.219 kappa and 0.643 ROC-AUC. Continuous regression remained weak, with best performance of R² = 0.052, RMSE = 121.7 Mg ha⁻¹, MAE = 96.1 Mg ha⁻¹ and Spearman’s ρ = 0.242.
Conclusions: The study demonstrates that exact four-class GEDI biomass discrimination is limited in this dense tropical forest context, even with extensive predictor combinations and modelling strategies. Binary biomass stratification provides a more separable target, whereas continuous AGBD prediction remains strongly constrained, highlighting the effects of biomass saturation, spatial heterogeneity and limited radar-optical sensitivity.