Accurate estimation of stratified forest canopy height is crucial for understanding forest dynamics and estimating carbon stocks. While remote sensing is widely used for regional canopy mapping, its accuracy is often limited in mountainous and dense forests, particularly for understory canopy layers, because of signal saturation and restricted sensor penetration through the overstory canopy. To address these issues, we design a novel deep learning model, which combines the self-attention mechanism and the Vision Transformer encoder–decoder architecture (ViT-ed) with conditional modeling to achieve precise prediction of stratified canopy height using multi-temporal Sentinel-1/2 and Global Ecosystem Dynamics Investigation (GEDI) data. This model structure is tailored to the ecological correlations between vertical forest layers and optimized with a hybrid loss function and achieves significant improvements over the encoder-only convolutional neural network baseline (CNN-e). Validated against field inventory data, our model achieves a correlation coefficient of 0.843 (root mean square error (RMSE) = 1.532 m) for overstory canopy height and 0.825 (RMSE = 0.377 m) for understory canopy height, markedly outperforming the CNN-e benchmarks (RMSE = 1.846 m and 0.451 m, respectively). These results demonstrate improved understory canopy height estimation relative to CNN-e in the study area. Based on freely available data, this method generates continuous canopy height maps for the study area and has potential to enhance carbon stock estimation across broader regions.