【摘要】 森林冠层高度作为森林垂直结构的关键参数，其精准估测在碳循环与森林地上生物量研究中发挥着不可或缺的作用。随着遥感技术的不断发展，多源遥感数据为大尺度森林监测中冠层高度估测提供了新的可能性。本研究以中国东北地区为研究区域，提出了一种结合随机森林RF(Random Forests)和经验贝叶斯克里金EBK(Empirical Bayesian Kriging)方法的模型（RF-EBK），用于区域尺度森林冠层高度的估测。该模型基于星载激光雷达ICESat-2提供的离散冠层高度数据、Landsat 8 OLI影像、航天飞机雷达地形测绘任务地形数据以及森林冠层覆盖数据，首先采用基于交叉验证的递归特征消除方法对多源遥感数据中提取的特征因子进行筛选；然后通过RF模型进行森林冠层高度估测，并计算测试集的估测残差；最后基于估测残差的空间自相关性，利用EBK方法对估测残差进行建模，得到研究区域空间连续残差插值结果，并对RF估测结果进行残差校正，从而有效提高模型的估测精度，最终实现中国东北地区2023年30 m森林冠层高度的高精度估测。结果表明，森林冠层覆盖特征因子在冠层高度估测中的重要性较高。在模型精度方面，RF-EBK模型相较于单独使用RF模型具有更优的估测性能，验证集R2提高了59.52%,RMSE和rRMSE均降低了27%。此外，使用在研究区域内6个采样点采集的无人机激光雷达数据对RF-EBK模型估测结果进行精度验证，R2为0.69,RMSE为1.65 m,rRMSE为7.81%。综上，RF-EBK模型能够实现区域尺度森林冠层高度的高精度估测，为中国东北地区的精准营林管理和可持续森林资源经营提供了有效的技术支持。更多还原

【Abstract】 Forest canopy height, a key parameter that reflects the vertical structure of forests, is essential for understanding the structure and function of forest ecosystems. Accurate estimation of canopy height is important for carbon cycle assessments, above-ground biomass estimation, and ecosystem health monitoring. With the continuous advancement of remote sensing technologies, particularly the integration of LiDAR and optical remote sensing data, the potential for estimating forest canopy height at regional scales has become increasingly prominent, making it a current research hotspot in forest resource monitoring. This study focuses on Northeast China(NEC) and proposes a hybrid model that integrates Random Forest(RF) and Empirical Bayesian Kriging(EBK), referred to as the RF-EBK model, to enhance the accuracy and robustness of regional-scale canopy height estimation. The model incorporates discrete canopy height data from spaceborne LiDAR ICESat-2(ATL08), Landsat 8 OLI imagery, Shuttle Radar Topography Mission elevation data, and forest canopy cover data. Initially, a recursive feature elimination method with cross validation is employed to select optimal variables, reduce redundancy, and improve the model’s generalization ability. The RF model is then used to produce initial canopy height estimates, and residuals are calculated using a test dataset. Given the spatial autocorrelation of the residuals, the EBK method is applied to spatially model and interpolate them, generating a continuous residual surface across the study area. This residual surface is used to correct the RF predictions, effectively improving estimation accuracy. Ultimately, a highly accurate forest canopy height map at 30 m resolution for NEC in 2023 is produced. Results show that forest canopy cover is the most important variable in the model. Among the topographic factors, slope, elevation, and aspect are highly influential, reflecting the remarkable role of terrain in vegetation type and growth conditions. In terms of optical remote sensing features, the original Landsat 8 OLI bands, namely, B2, B4, and B7, exhibit high importance. Moreover, texture features derived from bands B3, B6, and B7(i.e., B3＿savg, B6＿savg, and B7＿savg) are more important than those from the original bands, underscoring the value of incorporating spatial texture features into canopy height estimation. Tasseled cap greenness, indicative of canopy cover and vegetation health, also shows strong predictive power. In terms of model performance, the RF-EBK model considerably outperforms the standalone RF model by effectively mitigating the overestimation of low canopy heights and underestimation of high canopy heights. After residual correction, the coefficient of determination(R2) on the validation set increases by 59.52%, and the root mean square error(RMSE) and relative root mean square error(rRMSE) decrease by 27%. Furthermore, canopy height measurements extracted from unmanned aerial vehicle laser scanning data collected from six sites are used as reference data for model validation. Results show that the RF-EBK model achieves high accuracy, with an R2 of 0.69, RMSE of 1.65 m, and rRMSE of 7.81%. In conclusion, the RF-EBK model is a reliable approach for highly accurate estimation of forest canopy height at the regional scale and offers robust technical support for precision silviculture and sustainable forest resource management in NEC.更多还原