Species information in multi-temporal Sentinel-2 data improves forest canopy height estimation

• Species information largely explains CHM improvements from multi-temporal data. • Multi-temporal data improves CHM across broadleaf, needleleaf, and mixed forests. • Value of species maps for CHM peaks in species-mixed pixels, drops in pure stands. • CHM accuracy depends on number and timing of Sentinel-2 images. The integration of Sentinel-2 (S2) multispectral data and LiDAR measurements using convolutional neural networks (CNNs) has been extensively explored for large-scale canopy height mapping. Recent studies reported multi-temporal S2 imagery improves canopy height mapping, but the ecological mechanism underlying this improvement remains poorly understood. Here, we hypothesize that this improvement in canopy height estimation is primarily because multi-temporal data contain information relevant to tree species composition. Each tree species likely exhibits a unique reflectance–height relationship, suggesting that incorporating species information could improve canopy height estimates. We further hypothesize that simple plant functional type maps can improve canopy height estimation but less improvement than detailed species-level information. To test these hypotheses, we trained CNN models using airborne LiDAR canopy height at nine National Ecological Observatory Network sites including broadleaf, needleleaf, and mixed forests, using four input combinations: [1] single acquisition only, [2] multi-temporal data only, [3] single acquisition with species information, and [4] multi-temporal data with species information. The hypothesis is supported by our findings that first, multi-temporal data improved the MAE of canopy height estimation by approximately 21 % on average across sites compared to the models using single acquisitions ([1] vs [2]) and second, adding species information input reduced the MAE by around 15 % for single acquisition based models ([1] vs [3]), while adding species information to multi-temporal data resulted in only a marginal (5 %) reduction ([2] vs [4]). Furthermore, adding detailed species data to the models consistently provides better performance compared to adding plant functional type maps. Further experiments reveal that canopy height estimation accuracy depends on both the number and timing of S2 acquisitions. These findings clarify the mechanistic relationship between seasonal reflectance, species, and canopy structure, underscoring that incorporating species information is key to reducing estimation bias in tall forests for large-scale biomass assessments.