HSFFNet: A Height-Guided Spatial–Frequency Fusion Network for Crown-Level Tree Species Segmentation

High-resolution UAV RGB imagery and canopy height models (CHMs) can provide complementary appearance and structural information for tree crown analysis. However, existing multimodal segmentation methods often treat optical and height data as parallel feature sources, and the structural role of canopy height in tree crown representation remains insufficiently explored. To address this limitation, we propose HSFFNet, a Height-Guided Spatial–Frequency Fusion Network for crown-level tree species segmentation, which exploits CHM-derived structural information to enhance RGB representations across multiple feature scales and frequency components. Specifically, a Scale-Aligned Height Encoder (SAHE) is introduced to construct hierarchical height representations corresponding to the multilevel RGB features. A Height-Guided Spatial–Frequency Fusion (HGSFF) module then models appearance and structure interaction in the spatial domain while using canopy height to regulate frequency-specific RGB responses. Finally, a Multi-Scale Height-Frequency Refinement (MHFR) module is applied to extract localized height variations at multiple spatial ranges to complement structural details. Experiments on two tree species datasets show that HSFFNet achieves mIoU values of 78.26% and 79.05%, respectively, outperforming several representative RGB-based and multimodal segmentation methods.