M3FNet: Multi-modal multi-temporal multi-scale data fusion network for tree species composition mapping

Accurate estimation and mapping of t ree s pecies c omposition (TSC) is crucial for sustainable forest management. Recent advances in Light Detection and Ranging (lidar) technology and the availability of moderate spatial resolution, surface reflectance time series passive optical imagery offer scalable and efficient approaches for automated TSC estimation. In this research we develop a novel deep learning framework, M3F-Net (Multi-modal, Multi-temporal, and Multi-scale Fusion Network), that integrates multi-temporal Sentinel-2 (S2) imagery and single photon lidar (SPL) data to estimate TSC for nine common species across the 630,000-hectare Romeo Malette Forest in Ontario, Canada. A dual-level alignment strategy combines (i) superpixel-based spatial aggregation to reconcile mismatched resolutions between high-resolution SPL point clouds (>25 pts/m 2 ) and coarser S2 imagery (20 m), and (ii) a grid-based feature alignment that transforms unordered 3D point cloud features into structured 2D representations, enabling seamless integration of spectral and structural information. Within this aligned space, a multi-level Mamba-Fusion module jointly models multi-scale spatial patterns and seasonal dynamics through selective state-space modelling, efficiently capturing long-range dependencies while filtering redundant information. The framework achieves an R 2 score of 0.676, outperforming existing point cloud-based methods by 6% in TSC estimation. For leading species classification, our results are 6% better in terms of weighted F1, using either the TSC-based method or the standalone leading species classification method. Addition of seasonal S2 imagery added a 10% R 2 gain compared to the SPL-only mode. These results underscore the potential of fusing multi-modal and multi-temporal data with deep learning for scalable, high-accurate TSC estimation, offering a robust tool for large-scale management applications.