Mapping the Big Trees of Vancouver Island with LiDAR, Sentinel-1, Sentinel-2 and Deep Learning

Decades of overharvesting have transformed much of Vancouver Island’s productive temperate rainforests into young, homogeneous forests, making the few remaining large-tree forests (height >55m) both rare and valuable. To map and conserve these forests, we produce a high-resolution canopy height model using a UNET deep-learning approach trained on LiDAR-derived 99th percentile height returns (~29% of the study area) and Sentinel-1, -2, ALOS PALSAR, and geographical predictors. The best-performing model (R2=0.76, MAE=4.73 m, and BIAS= -0.86 m), trained with Sentinel-1(VVVH) and Sentinel-2 (RGBNIR), outperformed all global canopy height products. We found a total of 135,482 location with canopy height above 55 m. Large tree forests were disproportionally found on upper slopes (~55%, 1.2x higher than expected) and valleys (28%, > 4x higher), and underrepresented in mid-slopes (7.4% in mid-slope, ~4x lower) and plateaus (1.4% in plateaus, 7x lower). Over half (63%) of forests sustaining big trees remain unprotected. These rare forests hold immense value yet constitute only a small fraction of the forested landscape, underscoring the urgency of a targeted conservation strategy.