Retrieving Forest Canopy Height From ICESat-2 Data by an Improved DRAGANN Filtering Method and Canopy Top Photons Classification

Abstract

Lots of noise photons limit the application of Ice, Cloud and land Elevation Satellite-2 (ICESat-2) in forest canopy height retrieval. Abundant noise photons lead to unstable filtering results, thus affecting the canopy top photons classification. Therefore, this study proposes a method that takes account of the background noise level in the photon cloud. First, we propose an improved Differential, Regressive, and Gaussian Adaptive Nearest Neighbor (DRAGANN) filtering approach based on the DRAGANN filtering method. To obtain a more stable filtering result, large-scale and small-scale search radiuses are combined to improve the DRAGANN filtering performance. Second, we retrieve sub-canopy terrain topography by the multi-scale window detection method from the filtered photon cloud. Finally, a robust photon acquisition criterion based on the elevation difference of the filtered photons and the uneven density of signal photons is proposed to extract the canopy-top surface, which aims to mitigate the influence of inconsistencies of residual noise photons along the track. In addition, considering the fluctuation of the canopy surface, we use a non-spline interpolation method to obtain a continuous canopy surface, which avoids the Runge phenomenon caused by the spline interpolation method. The ICESat-2 data acquired in the Harvard Forest Region (HARV) is selected to assess the proposed method. The filtering performance of the improved DRAGANN approach shows stable than that of the DRAGANN approach. The proposed method’s root means square error (RMSE) and coefficient of determination (R2) reach 3.85 m and 0.55, respectively. The results indicate that the proposed method can retrieve reliable forest canopy height from the ICESat-2 data and performs significantly better than the ATL08 canopy height product in the test site.