Forest-type-aware canopy height estimations using multi-source remote sensing data by integrating terrain correction mechanisms

Forest canopy height (FCH) is a critical parameter for characterizing forest structure and carbon dynamics. However, large-scale FCH estimation using ICESat-2 faces challenges in terrain-induced errors and neglect of forest-type-specific structural variability. To address these issues, this study proposes a forest-type-aware, fine-resolution inversion method that integrates multi-source remote sensing data. Firstly, this study utilizes a 1 m-resolution digital terrain model (DTM) generated from airborne LiDAR data provided by the National Ecological Observatory Network (NEON) to perform terrain correction on ICESat-2 canopy photons, significantly reducing errors caused by terrain variability. Additionally, vertical datum correction is applied using the Geoid12A geoid model to ensure accurate height measurements. Secondly, the study extracts a comprehensive set of 33 features, including spectral, radar, topographic, and climatic variables, is extracted from Sentinel-1/2, SRTM, and WorldClim datasets. To optimize model performance and reduce computational complexity, the Random Forest Recursive Feature Elimination (RF-RFE) algorithm is employed to dynamically identify the most informative feature subsets for different forest types (deciduous, evergreen, and mixed). Thirdly, forest-type-specific canopy height models are developed using Random Forest, with model parameters optimized through a cross-validation strategy. The Great Smoky Mountains National Park (USA) is selected as the study area, with NEON-provided canopy height models serving as reference data. Experimental results demonstrate that the proposed method achieves reliable prediction accuracy, with an overall R 2 of 0.714 and RMSE of 5.070 m. Notably, the integration of high-precision DTM for terrain correction significantly improves estimation accuracy, with R 2 increasing to 0.809 and RMSE decreasing to 3.602 m. Furthermore, the forest-type-specific modeling framework outperforms a unified model, achieving higher R 2 and more stable performance across diverse forest types. This study presents an innovative multi-source data fusion framework that integrates terrain correction mechanisms with forest type classification, thereby improving the accuracy and robustness of large-scale FCH estimation. The proposed method offers substantial potential for application in global carbon cycle analysis, biomass estimation, and climate change research. • Developed a forest type-aware multi-source fusion method. • Reduced terrain-induced errors by introducing high-precision DTM. • Optimized features via RF-RFE, boosting R 2 and reducing MAE across forest types. • Achieved robust adaptation to structural heterogeneity by type-specific modeling.