Satellite hyperspectral imagery (HSI) fused with LiDAR enables large-area, fine-grained tree-species mapping in subtropical forests. However, current satellite HSI-LiDAR fusion remains limited because non-task-oriented spectral reduction may weaken discriminative narrowband and inter-band cues, while weakly coupled fusion cannot adequately address cross-modal inconsistency in satellite observations. To address these issues, we propose MUFNet, a mamba-based uncertainty-aware fusion framework that integrates multi-stage cross-modal interactions with uncertainty-guided prediction aggregation under end-to-end task-driven representation learning. We validate MUFNet in a representative subtropical region and benchmark it against six state-of-the-art (SOTA) models. Experimental results demonstrate that MUFNet achieves the best mapping performance, with an overall accuracy (kappa) of 93.34% (91.17%). Moreover, MUFNet significantly outperforms all competing methods, improving overall accuracy (kappa) by 2.85%–11.45% (6.45%–15.49%). In addition, MUFNet exhibits good predictive reliability, achieving an ECE of 0.0073, and maintains a favorable performance–efficiency trade-off with only 0.400 M parameters and 0.086 GFLOPs. Ablation studies further confirm the complementarity of the three fusion stages and the contributions of the proposed task-driven representation learning, cross-modal interaction, and uncertainty-aware routing components. This work provides a transferable technical pathway for fine-grained tree species mapping under multi-source remote sensing, offering methodological support for biodiversity conservation and forest resource management.

Accurate tree species mapping via satellite imagery is essential for forest resource management at both global and regional scales. Forest phenological processes, captured in Sentinel-2 satellite image time series (SITS), generate distinct spatial and spectral features across different stages. However, previous studies primarily focused on temporal features and sequential relationships, often overlooking pixel-level distortions and spatial shifts at canopy boundaries, which hinder accurate mapping in transition zones between tree species. In this study, we propose the optical flow alignment and physical consistency constraint network (OAPCNet), a novel deep learning approach designed to track pixel-level spatial drift and dynamically correct feature distortions in SITS data. By incorporating a physical consistency loss, OAPCNet reduces alignment errors and physical distortions. We applied OAPCNet to map 16 tree species across three distinct temperate climate zones in the Northern Hemisphere—temperate monsoon, oceanic, and continental—and compared its performance with six state-of-the-art (SOTA) models. Results show that OAPCNet achieved the highest mapping accuracy, with overall accuracy (kappa) ranging from 89.73% to 91.12% (86.17% to 87.93%) and low standard deviation (<0.9%) across different regions. Furthermore, it outperformed SOTA models, improving overall accuracy (kappa) by 1.92%-9.29% (2.76%-12.63%). Ablation experiments demonstrate the critical contribution of optical flow alignment with constraints and the temporal feature enhancement-fusion strategy to robust temporal feature representation. This study introduces a robust deep learning approach for global temperate forest species mapping, crucial for biodiversity monitoring and phenology assessment using Sentinel-2 imagery.

Accurately identifying dominant forest tree species from remote sensing supports biodiversity assessment, carbon stock estimation, and sustainable forest management. Sentinel-2 satellite time series capture phenological dynamics and canopy spectral variation, providing valuable information for regional tree species classification. However, interannual shifts in phenological timing can reduce the temporal correspondence among repeated annual cycles. This temporal misalignment complicates the consistent representation of multiyear observations. To address this problem, this study proposes the Hierarchical Season-Year Representation Alignment Network (HSRA-Net). Using multiyear Sentinel-2 NDVI time series, HSRA-Net builds robust temporal representations, aligns comparable seasonal phases across years, and adaptively integrates informative temporal features. The model was evaluated on dominant tree species in representative temperate and subtropical forests in China using ground reference observations and four temporal scales: intra-growing, annual, biennial, and triennial. It was compared with standard models and representative recent spatiotemporal methods. Among the four temporal scales, HSRA-Net achieved its highest classification accuracy in the triennial scale in both study areas. HSRA-Net achieved OA of 88.90% and 83.17% in the temperate and subtropical regions, respectively. Compared with the intra-growing season, the triennial scale increased OA by 3.72% and 14.31%, respectively. These findings demonstrate that organizing comparable phenological information across repeated annual cycles can improve dominant tree species classification and highlight the effectiveness of HSRA-Net for multiyear temporal representation.