Integrating multimodal data through a data-driven ensemble model for optimizing forest height estimation

Accurate forest-height estimation is essential for ecological monitoring and forest management. This study applies an adaptive stacking-based fusion framework that screens combinations of out-of-fold predictions from three regression models and three existing canopy-height products within the training data. Applied to the Guangdong–Hong Kong–Macao Greater Bay Area, nested random five-fold cross-validation yielded r = 0.54, RMSE = 7.54 m and MAE = 5.61 m. Stacking had the lowest observed pooled RMSE, although its margin over RF and GBDT was small and it was not consistently best across height classes. Buffered spatial block cross-validation yielded r = 0.53, RMSE = 7.58 m and MAE = 5.66 m. Cross-fitted residual scaling gave 90.0% empirical pooled coverage for nominal 90% intervals, but only 45.3% for RH98 ≥ 30 m. Because Global Ecosystem Dynamics Investigation (GEDI) RH98 was both the target and validation reference and all three external products incorporate GEDI observations, these metrics quantify agreement with withheld GEDI RH98 rather than independently verified absolute canopy-height accuracy.