Jie Tang

Aboveground biomass (AGB) serves as an important indicator for assessing the productivity of forest ecosystems and exploring the global carbon cycle. However, accurate estimation of forest AGB remains a significant challenge, especially when integrating multi-source remote sensing data, and the effects of different feature combinations for AGB estimation results are unclear. In this study, we proposed a method for estimating forest AGB by combining Gao Fen 7 (GF-7) stereo imagery with data from Sentinel-1 (S1), Sentinel-2 (S2), and the Advanced Land Observing Satellite digital elevation model (ALOS DEM), and field survey data. The continuous tree height (TH) feature was derived using GF-7 stereo imagery and the ALOS DEM. Spectral features were extracted from S1 and S2, and topographic features were extracted from the ALOS DEM. Using these features, 15 feature combinations were constructed. The recursive feature elimination (RFE) method was used to optimize each feature combination, which was then input into the extreme gradient boosting (XGBoost) model for AGB estimation. Different combinations of features used to estimate forest AGB were compared. The best model was selected for mapping AGB distribution at 30 m resolution. The outcomes showed that the forest AGB model was composed of 13 features, including TH, topographic, and spectral features extracted from S1 and S2 data. This model achieved the best prediction performance, with a determination coefficient (R²) of 0.71 and a root mean square error (RMSE) of 18.11 Mg/ha. TH was found to be the most important predictive feature, followed by S2 optical features, topographic features, and S1 radar features.

In complex mountainous terrain, the terrain slope causes the scattering of pulsed lasers, generating a lot of noise in photon cloud data (PCD) collected from forestland, which seriously affects the accurate retrieval of forest structure parameters. To address this problem, a multilevel auto-adaptive denoising (MLAD) algorithm was proposed in this study. Firstly, random noise photons (NPs) were removed through the ordering points to identify the clustering structure (OPTICS) algorithm in the coarse denoising process. Secondly, in the fine denoising step, the circular search domain in the OPTICS algorithm was replaced with an elliptical search domain. The photons after coarse denoising were automatically divided along-track direction into several continuous segments of 100 m each. The median slope method was used to automatically calculate the slope of the forested terrain in each interval segment, so that the range of the ellipse search domain was automatically adjusted to achieve accurate denoising of PCD. Finally, the denoising results of the MLAD algorithm in three different forested terrain areas were compared with those of the Difference, Regression, and Gaussian Adaptive Nearest Neighbor (DRAGANN) algorithms, and the performance of the MLAD algorithm was evaluated for both different terrain slopes and different vegetation coverages. The results indicated that compared with the DRAGANN algorithm, the MLAD algorithm has higher denoising capability in different regions. The denoising results of the MLAD algorithm exhibit slight changes with the variation in slope, and the F-values are around 0.96, demonstrating good robustness. The F-value of the MLAD algorithm mostly exceeds 0.95 in different vegetation coverages. Overall, the MLAD algorithm exhibits stronger noise identification capabilities for complex forest environments. These results can provide a reference for subsequent accurate extraction of forest structural parameters.