Cross-Causal CarbonFormer: Rethinking Forest Carbon Density Estimation With a Causal Spatiotemporal Framework

Unofficial AI-generated summary based on the public title and abstract. Not an official translation.

🔬研究者:Offers a novel causal spatiotemporal approach for carbon density estimation that outperforms standard CNNs and generalizes across biomes without retraining.

🏢実務担当者:Can be used for high-accuracy forest carbon mapping and monitoring, especially for fragmented landscapes and trees outside forests, supporting carbon project validation.

🏛政策担当者:Provides evidence that trees outside forests contribute nearly half of landscape carbon, informing more inclusive carbon accounting in national inventories and climate mitigation strategies.

Forest aboveground carbon density (ACD) estimation is essential for characterizing ecosystem carbon cycles and supporting climate mitigation strategies. Due to their excellent local contextual modeling ability, convolutional neural networks (CNNs) have emerged as powerful feature extractors for remote sensing-based carbon mapping. However, CNNs often fail to represent the complex causal sequence attributes of multiyear carbon trajectories and the structural heterogeneity of fragmented landscapes due to inherent backbone limitations. To this end, we rethink ACD estimation from a causal spatiotemporal perspective and propose CarbonFormer. CarbonFormer also employs a stacked long short-term memory-based temporal causal decoder that explicitly preserves causal temporal ordering; a physically meaningful constraint for sequential vegetation dynamics spanning decadal observation periods (2015–2024). Beyond static biophysical representations, CarbonFormer jointly learns long-term vegetation dynamics through a fusion attention module that adaptively fuses multispectral, topographic, and Global Ecosystem Dynamics Investigation LiDAR-derived structural metrics. To enforce ecological realism, we formulate a novel Cross-Causal Loss Function jointly optimizing spatial accuracy, interannual trajectory consistency, and a physical carbon growth constraint; Extensive experiments demonstrate superiority over standard baselines methods. To assess generalization, we evaluate CarbonFormer on three held-out forest regions: Amazon Basin, Brazil; Lopé NP, Gabon; and Siberia, Russia, confirming cross-biome transferability without target-domain retraining. Further analysis reveals that dense forests account for 51.4% of total carbon while trees outside forests contribute 48.6% to the landscape-scale carbon stock.

• semanticscholar https://doi.org/10.1109/jstars.2026.3695334first seen 2026-07-24 06:37:43