Accurately localizing individual trees and identifying species distribution are critical tasks in forestry remote sensing. Visual Simultaneous Localization and Mapping (visual SLAM) algorithms serve as important tools for outdoor spatial positioning and mapping, mitigating signal loss caused by tree canopy obstructions. To address these challenges, a semantic SLAM algorithm called LPD-SLAM (Line-Point-Distance Semantic SLAM) is proposed, which integrates stereo cameras with an inertial measurement unit (IMU), with contributions including dynamic feature removal, an individual tree data structure, and semantic point distance constraints. LPD-SLAM is capable of performing individual tree localization and tree species discrimination tasks in forest environments. In mapping, LPD-SLAM reduces false species detection and filters dynamic objects by leveraging a deep learning model and a novel individual tree data structure. In optimization, LPD-SLAM incorporates point and line feature reprojection error constraints along with semantic point distance constraints, which improve robustness and accuracy by introducing additional geometric constraints. Due to the lack of publicly available forest datasets, we choose to validate the proposed algorithm on eight experimental plots, which are selected to cover different seasons, various tree species, and different data collection paths, ensuring the dataset’s diversity and representativeness. The experimental results indicate that the average root mean square error (RMSE) of the trajectories of LPD-SLAM is reduced by up to 81.2% compared with leading algorithms. Meanwhile, the mean absolute error (MAE) of LPD-SLAM in tree localization is 0.24 m, which verifies its excellent performance in forest environments.

The Terrestrial Ecosystem Carbon Inventory Satellite (TECIS) provides highly dense full waveform LiDAR data for global carbon monitoring. However, in forest environments, the complex canopy structure and terrain introduce significant background noise, making effective signal extraction challenging. To address this issue, an adaptive waveform denoising framework is proposed in this study. Bayesian optimization (BO) is first employed to adaptively determine the effective signal threshold for each waveform by minimizing structural volatility within background noise segments, thereby achieving reliable separation between effective signals and background noise. Subsequently, a structure-aware adaptive Gaussian filtering (SAAF) method is developed. Residual noise in the effective signal is suppressed by dynamically adjusting the Gaussian kernel width based on structural factors, including notch density, flat peak characteristics, peak skewness, and multipeak structures. The results indicate that, compared to the official TECIS parameter extraction (TPE) method, the BO method removes an additional 5.14% of redundant background noise energy, yielding a cleaner, more physically realistic effective signal. Furthermore, relative to manually annotated effective signal start and end points, the localization accuracy of the extracted start and end points is improved by 77.5% and 78.3%, respectively. For effective signal smoothing, SAAF maintains high precision across varying slopes and forest types, with peak signal-to-noise ratio (PSNR) consistently exceeding 44 dB and the correlation coefficient (CC) remaining above 0.96. This framework provides reliable, structurally complete waveform data for subsequent inversion of forest structure parameters.