Using airborne lidar and machine learning to predict visibility across diverse vegetation and terrain conditions
Using airborne lidar and machine learning to predict visibility across diverse vegetation and terrain conditions
复制标题
利用机载激光雷达和机器学习来预测不同植被和地形条件下的能见度
DOI：
10.1080/13658816.2023.2224421
复制
发表时间：
2023-07
影响因子：
5.7
通讯作者：
K. Mistick;Michael J. Campbell;Matthew P. Thompson;P. Dennison
中科院分区：
文献类型：
--
作者：
K. Mistick;Michael J. Campbell;Matthew P. Thompson;P. Dennison
摘要 在许多学科中使用的可见性分析依赖于视域算法，该算法根据给定的表面模型映射观察者可见的位置。由于计算成本极高，在大范围内绘制连续可见性地图并不常见。这项研究引入了一种使用机载激光雷达和随机森林进行空间详尽可见性映射的新方法，该方法仅需要稀疏的视域样本。在美国本土 24 个地形和植被多样化的景观中，使用 1 m 分辨率激光雷达衍生的数字表面模型，在四个不同的观测半径（125 m、250 m、500 m、1000 m）处生成了 1000 个随机点视域。能见度指数（可见面积占总面积的比例）被用作场地尺度和国家尺度建模的目标变量，该建模使用一组不同的 146 个基于地形和植被的 10 m 分辨率指标作为预测因子。基于植被的变量，尤其是基于当地社区的变量，比基于地形的变量更重要。更准确地估计了较短距离的能见度。在更广泛的植被和地形条件上训练的国家规模模型导致 R2 得到改善，尽管在某些地点误差与地点规模模型相比有所增加。独立测试地点的结果证明了该方法在不同景观中的应用潜力。
Abstract Visibility analyses, used in many disciplines, rely on viewshed algorithms that map locations visible to an observer based on a given surface model. Mapping continuous visibility over broad extents is uncommon due to extreme computational expense. This study introduces a novel method for spatially-exhaustive visibility mapping using airborne lidar and random forests that requires only a sparse sample of viewsheds. In 24 topographically and vegetatively diverse landscapes across the contiguous US, 1000 random point viewsheds were generated at four different observation radii (125 m, 250 m, 500 m, 1000 m), using a 1 m resolution lidar-derived digital surface model. Visibility index – the proportion of visible area to total area – was used as the target variable for site-scale and national-scale modeling, which used a diverse set of 146 terrain- and vegetation-based 10 m resolution metrics as predictors. Variables based on vegetation, especially those based on local neighborhoods, were more important than those based on terrain. Visibility at shorter distances was more accurately estimated. National-scale models trained on a wider range of vegetation and terrain conditions resulted in improved R2, although at some sites error increased compared to site-scale models. Results from an independent test site demonstrate potential for application of this methodology to diverse landscapes.