Deep learning-driven forest canopy height mapping in boreal regions through multi-source remote sensing fusion: Integrating Sentinel-1/2, PALSAR, and ICESat-2/LVIS data

Forest canopy height is a key indicator for estimating forest carbon sinks and managing vegetation growth. Existing methods for fusing optical and Light Detection and Ranging (LiDAR) data still have limitations in canopy height estimation for boreal forests. In this study, we develop a forest canopy height estimation model (VGG-AdaBins) that leverages convolutional neural networks (CNNs) to extract deep features from multi-source remote sensing data. By introducing an adaptive tree height distribution estimation module, the model enables the coupling of multi-source remote sensing data for forest canopy height estimation. A joint validation dataset is constructed, including Sentinel-1/2, PALSAR images, airborne LVIS LiDAR, and spaceborne ICESat-2 photon-counting LiDAR data. This dataset is used to train the canopy height model. Finally, the performance of the canopy height prediction model is evaluated using 100 independent airborne datasets. The model’s prediction of tree height shows an MAE of 1.42 m and a RMSE of 2.25 m. The predicted 30 m canopy height map exhibits good consistency with the existing airborne data and demonstrated higher accuracy compared with current forest canopy height maps, with an accuracy improvement of at least 20%. The high prediction accuracy demonstrates that VGG-AdaBins, by integrating multi-source remote sensing data, can map continuous canopy height at the regional scale. This approach contributes to the advancement of large-scale canopy height mapping and forest carbon stock assessment in boreal forests.