Answers are grounded in this paper and cite their sources.
Accurate estimation of forest aboveground biomass (AGB) is a critical prerequisite for regional carbon accounting. However, existing direct regression methods lack synergistic constraints on the coupling relationships among forest structural variables, such as tree height and diameter at breast height, often causing biomass predictions in complex stands to become homogenized, with systematic underestimation of high values, overestimation of low values, and blurred stand boundaries. To address this, we reframe forest AGB retrieval as a structural variable-driven conditioned mapping problem and propose a hybrid expert attention multitask model (HEAMM). The model employs gated shifted-window self-attention to capture spatial context relevant to forest structure, leverages sparse expert routing in a mixture-of-experts feedforward network (MoE-FFN) to enhance nonlinear mapping capability under diverse local remote-sensing-structure patterns, and constructs an allometric scaling soft constraint based on Repola’s species-specific allometric equations, providing a probabilistic ecological prior to guide the joint prediction of structural variables and AGB estimation. On the Finnish Forest Centre dataset, we established five experimental scenarios: airborne hyperspectral data from TAIGA (0.6 m) and ESSD (0.5 m), AEFORGE-TAIGA and AEFORGE-ESSD based on 10 m spaceborne AlphaEarth embeddings, and RJ-Transfer (a spatially independent test area transfer). HEAMM achieved AGB R² values of 0.915, 0.978, 0.952, 0.961, and 0.824, respectively. Ablation experiments indicate that removing the MoE-FFN led to the most pronounced decline in AGB estimation performance ( -5.1% ); the gain from the allometric scaling soft constraint (AEC) ( +5.9% ) exceeded that from incorporating canopy height model structural information ( +3.7% ), and their combined gain was nearly additive ( +7.3% ). Spatial predictions and scatter analysis further reveal that HEAMM effectively alleviates the homogenized outputs typical of direct AGB prediction, producing sharper stand boundaries and more pronounced structural gradients and high–low biomass stratification. The results suggest that joint estimation of structural variables, guided by an allometric probabilistic ecological prior, can simultaneously enhance the numerical accuracy and spatial representation quality of stand-scale AGB retrieval, deliver forest structural parameters and biomass products for complex stands, and provide reliable support for forest carbon stock assessment.
No takes yet. Share an insight, caveat, or question.
Hao et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: