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

Forest aboveground carbon density (ACD) estimation is essential for characterizing ecosystem health 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 carbon trajectories and the structural heterogeneity of fragmented landscapes due to the limitations of their inherent network backbones. To solve this issue, we rethink ACD estimation from a causal spatiotemporal perspective and propose a novel backbone framework called CarbonFormer. Beyond static biophysical representations, CarbonFormer is capable of learning long-term vegetation dynamics through a Temporal Causal Decoder, yielding groupwise temporal embeddings from decadal observation sequences (2015–2024). More significantly, to capture high-order interactions across disparate modalities, we devise a Fusion Attention Module that adaptively learns to fuse “soft” residuals from multispectral, topographic, and LiDAR-derived structural metrics. It is worth noting that the proposed CarbonFormer is a highly flexible backbone network, applicable to both pixel-level estimation and landscape-scale mapping. We evaluate the performance of CarbonFormer across diverse forest ecosystems by conducting extensive experiments, demonstrating superiority over standard sequence models and outperforming standard baselines. Further analysis reveals that dense forests account for 51.4% of total carbon, while trees outside forests contribute a significant 48.6% to the landscape-scale carbon stock. Codebase will be made available at GitHub.