Generating Wall-to-Wall Canopy Height Information from Discrete Data Provided by Spaceborne LiDAR System

Abstract

Provision of multi-temporal wall-to-wall canopy height information is one of the initiatives to combat deforestation and is necessary in strategizing forest conversion and reforestation initiatives. This study generated wall-to-wall canopy height information of the subtropical forest of Lishan, Taiwan, using discrete data provided by spaceborne LiDARs, wall-to-wall passive and active remote sensing imageries, topographic data, and machine learning (ML) regression models such as gradient boosting (GB), k-nearest neighbor (k-NN), and random forest (RF). ICESat-2- and GEDI-based canopy height data were used as training data, and medium-resolution passive satellite image (Sentinel-2) data, active remote sensing data such as synthetic aperture radar (SAR), and topographic data were used as regressors. The ALS-based canopy height was used to validate the models’ performance using root mean square error (RMSE) and percentage RMSE (PRMSE) as validation criteria. Notably, GB displayed the highest accuracy among the regression models, followed by k-NN and then RF. Using the GEDI-based canopy height as training data, the GB model can achieve optimum accuracy with an RMSE/PRMSE of 8.00 m/31.59%, k-NN can achieve an RMSE/PRMSE of as low as 8.05 m/31.78%, and RF can achieve optimum RMSE/PRMSE of 8.16 m/32.24%. If using ICESat-2 data, GB can have an optimum RMSE/PRMSE of 13.89 m/54.86%; k-NN can have an optimum RMSE/PRMSE of 14.32 m/56.56%, while RF can achieve an RMSE/PRMSE of 14.72 m/58.14%. Additionally, integrating Sentinel-1 with Sentinel-2 data improves the accuracy of canopy height modeling. Finally, the study underlined the crucial relevance of correct canopy height estimation for sustainable forest management, as well as the potential ramifications of poor-quality projections on a variety of biological and environmental factors.