Development of a forest canopy height estimation model using GLAS full waveform data over sloping terrain

Abstract

ABSTRACTThe accurate estimation of forest canopy height is important because it leads to increased accuracy in the estimation of biomass, which is used in the study of the global carbon cycle, forest productivity, and climate change. However, there is no well-developed model that accurately estimates canopy height over undulating land. This paper describes the development of a back-propagation (BP) neural network model that estimates forest canopy height more accurately than other types of model. For modeling purposes, the land in the study area was classified as either plain (low relief areas) or hilly (high relief areas). Four different slope partition thresholds (5°, 10°, 15°, and 20°) were tested to determine the most suitable boundary value. ICESat-GLAS data provided by the Geoscience Laser Altimeter System (GLAS) aboard the Ice, Cloud and Land Elevation Satellite (ICESat), field survey data, and digital elevation model (DEM) data were collected and refined, and various parameters, including waveform extent and topographic index, were calculated. A BP neural network model was created to estimate forest canopy height. Two other models were also developed, one using the topographic index and the other using multiple linear regression, for comparison with the BP neural network model. After calibration, the three models were tested to assess the accuracy of the estimates. The results showed that the BP model estimated canopy height more accurately than the other two models. The use of a 10° boundary to partition the topography into low relief areas and high relief areas improved the accuracy of each model; using the 10° slope boundary, the coefficient of correlation r between the estimates given by the BP neural network model and the field-measured data increased from 0.89 to 0.95 and the Root Mean Square Error (RMSE) decreased from 1.01 to 0.73 m.