激光雷达作为一种主动遥感技术，凭借激光脉冲对森林良好的穿透性，可以对森林三维结构进行直接测量，能够快速获取高精度的森林空间结构和林下地形信息，在森林垂直结构观测中具有被动光学遥感无可比拟的优势。2018年9月，美国国家航空航天局（National Aeronautics and Space Administration，NASA）发射了冰，云和陆地高程卫星-2（Ice，Cloud，and land Elevation Satellite -2，ICESat-2）卫星，其搭载的ATLAS（Advanced Topographic Laser Altimeter System）激光雷达系统采用了全新的光子计数技术，正持续采集地球表面的高程信息，为大尺度林业监测提供了有效的数据源。光子计数激光雷达采样频率高，工作功率低，能够实现单光子水平的信号检测，是未来星载激光雷达载荷的重要选择之一。由于光子计数激光雷达对信号极为敏感，其接收的数据易受噪声干扰，因此，对光子数据进行有效去噪是后续科学研究与应用的重要前提。

本文以ICESat-2光子点云数据和高精度机载激光雷达数据为主要数据源，开展林区光子数据去噪方法研究和验证工作，主要内容有：

（1）针对目前已有光子去噪算法受地形影响严重的问题，提出一种基于旋转搜索区域的光子数据去噪方法。首先采用水平椭圆搜索区域的局部离群因子算法提取地面光子，并基于地面光子的高程对光子进行筛选。然后，利用预分类的地面光子计算地形坡度，旋转椭圆搜索区域，使其长轴与地面平行，计算光子的局部离群因子得分，设定去噪阈值，进而识别信号光子与噪声光子。之后，信号光子进一步被分类为冠层顶部光子、冠层光子和地面光子。本研究所提方法估算的地面高度相对机载数据（平均地面高度为1120米）的决定系数（R²）为1.00，平均绝对误差（Mean Absolute Error，MAE）为1.45米，均方根误差（Root Mean Square Error，RMSE）为2.82米。在冠层高度验证中，最佳研究尺度（80米）下估算的冠层高度与机载数据（平均冠层高度为12.3米）的R²、MAE和RMSE分别为0.86、1.82米和2.72米。上述结果表明，该算法能有效地对光子数据进行去噪，利用去噪光子估算的地形高度和冠层高度与机载激光雷达数据存在较好的一致性，同时冠层高度的估计误差，优于基于ICESat-2官方产品ATL08中去噪光子估计的冠层高度，表明该方法能够有效提升算法在复杂地形的去噪精度。

（2）基于旋转搜索区域的光子数据去噪方法能够在复杂地形情况下准确识别噪声光子，但其仍存在参数设定依赖经验的问题，为此，本研究提出一种基于XGBoost（eXtreme Gradient Boosting）的光子去噪方法。首先，参考机载激光雷达数据标注光子数据标签，提取光子的密度特征，构建光子数据集。其次，建立机器学习算法XGBoost模型，实现光子数据去噪。该方法去噪精度高，噪声预测正确率高于0.98，去噪效果优于ICESat-2官方产品ATL08，且基于该方法提取的冠层高度与机载激光雷达数据存在较好的一致性（R²、MAE和RMSE分别为0.97，0.63米和0.85米）。该方法以监督学习的方式解决了密度特征参数设定依赖经验的问题，能够实现光子数据快速去噪。在区域尺度内，该方法具有可迁移性，能够基于部分标记数据实现区域尺度的光子去噪任务。

LiDAR (Light Detection And Ranging), as an active remote sensing technology, can directly measure the three-dimensional structure of forests by virtue of the good penetration of laser pulses into forests, and can quickly obtain high-precision information on the spatial structure of forests and forest understory topography, which has unparalleled advantages in passive optical remote sensing in the observation of forest vertical structure.In September 2018, the National Aeronautics and Space Administration (NASA) launched the Ice, Cloud, and Land Elevation Satellite-2 (ICESat-2), which carries the ATLAS (Advanced Topographic Laser Altimeter System) lidar system, which uses a new photon-counting technology, is continuously collecting elevation information from the Earth's surface, providing an effective data source for large-scale forestry monitoring. The photon-counting LiDAR with its high sampling frequency and low operating power, enables signal detection at the single photon level and is one of the key options for future satellite-based LiDAR payloads. As photon-counting LiDAR is extremely sensitive to echo signals and its received data is susceptible to noise interference, effective denoising of photon data is an important prerequisite for subsequent scientific research and applications.

In this paper, ICESat-2 photon point cloud data and high-precision airborne LiDAR data are used as the main data sources to carry out study and validation of photon data denoising methods, with the following main contents:

(1) To address the problem that existing photon denoising algorithms are seriously affected by terrain, a photon data denoising method based on a rotating search region is proposed. Firstly, a local outlier factor algorithm for the horizontal ellipse search area is used to extract ground photons, and the photons are filtered based on the ground photon's elevation. Then, the pre-classified ground photons are used to calculate the terrain slope, rotate the ellipse search area so that its long axis is parallel to the ground, calculate the local outlier score of the photons, set the denoising threshold, and thus identify signal and noise photons. Afterwards, the signal photons are further classified into top-of-canopy photons, canopy photons and ground photons. The method proposed in this study estimates a coefficient of determination (R²) of 1.00 for terrain height relative to the airborne data (mean ground height of 1120 m), a MAE (Mean Absolute Error) of 1.45 m and a RMSE (Root Mean Square Error) of 2.82 m. In the canopy height validation, the R², MAE and RMSE of the estimated canopy height at the optimum study scale (80 m) compared to the airborne data (mean canopy height of 12.3m) were 0.86, 1.82m and 2.72m respectively. The above results show that the algorithm can effectively denoise the photons, and the terrain height and canopy height estimated using the denoised photons are in good agreement with the airborne LiDAR data, while the error of the estimated canopy height is better than the canopy height estimated based on the denoised photons in the ICESat-2’s official product ATL08, indicating that the method can effectively improve the denoising accuracy of the algorithm in complex terrain.

(2) The photon data denoising method based on rotating search area can accurately identify noisy photons in complex terrain, but it still suffers from the problem of experience-dependent parameter setting. To this end, a photon denoising method based on XGBoost (eXtreme Gradient Boosting) is proposed in this study. Firstly, the photon data is labelled with reference to the airborne LiDAR data, and the density features of the photons are extracted to construct a photon dataset. Secondly, the XGBoost (a machine learning algorithm) model is built to denoise the photon data. The method has high denoising accuracy and the accuracy of noise prediction is higher than 0.98. The denoising effect is better than the official ICESat-2 product ATL08, and the canopy height extracted based on this method is in good agreement with the airborne LiDAR data (R², MAE and RMSE are 0.97, 0.63 m and 0.85 m respectively). The method solves the problem of empirical dependence of density feature parameter setting in a supervised learning manner and enables fast denoising of photon data. At the regional scale, the method is portable and can achieve regional scale photon denoising tasks based on partially labelled data.