Forest Canopy Height Estimation Based on TECIS Multi-Angle Data and a Novel Discrete CHM Dataset Derived from GF-7 and GEDI

Forest canopy height (FCH) is a key parameter for estimating forest aboveground biomass (AGB) and carbon storage. The integration of GEDI relative height products and single-angle imagery has been widely applied in FCH estimation. However, GEDI geolocation errors and the canopy underestimation of traditional single-angle imagery in dense forests limit FCH estimation accuracy. To mitigate these limitations, this study proposes an FCH estimation method that integrates GaoFen-7 (GF-7) stereo pairs, GEDI ground elevation, and Terrestrial Ecosystem Carbon Inventory Satellite (TECIS) multi-angle data. First, this study integrated GF-7 stereo pairs and GEDI elevation to generate a discrete Canopy Height Model (CHM) dataset, providing reliable training samples for FCH estimation. Secondly, to mitigate the canopy underestimation issues associated with traditional single-angle features in FCH retrieval, we further introduced TECIS multi-angle data. To fully exploit its potential for structural characterization, we combined TECIS multi-angle features and topographic features with the discrete CHM dataset, utilizing a Bayesian-optimized LightGBM model to estimate FCH. The results show that: (1) The discrete CHM dataset constructed in this study (R2 = 0.65, RMSE = 2.39 m) outperformed the GEDI RH95 product without geolocation correction (R2 = 0.37, RMSE = 3.24 m). (2) After incorporating TECIS multi-angle features into the FCH inversion model, the predictive accuracy was improved, with the coefficient of determination R2 increasing from 0.36 to 0.51, and the RMSE and MAE decreasing by 0.33 m and 0.31 m. Meanwhile, the mean bias (MB) of trees above 15 m decreased from −2.12 m to −1.49 m, and the relative bias (rBias) decreased from −12.41% to −8.78%. (3) Independent validations of the estimation results using ALS-derived CHM (R2 = 0.46, RMSE = 2.40 m) and field measurements (R2 = 0.50, RMSE = 2.95 m) confirmed the overall predictive capability of the model.