Mixtures of airborne lidar-based approaches improve predictions of forest structure

Abstract

The most common method for modeling forest attributes with airborne lidar, the area-based approach, involves summarizing the point cloud of individual plots and relating this to attributes of interest. Tree- and voxel-based approaches have been considered as alternatives to the area-based approach but are rarely considered in an area-based context. We estimated three forest attributes: basal area, overstory biomass, and volume, across 1,680 field plots in Arizona and New Mexico. Variables from the three lidar approaches (area, tree, voxel) were created for each plot. Random forests were estimated using subsets of variables based on each individual lidar approach and mixtures of each approach. Boruta feature selection was performed on variable subsets, including the mixture of all lidar-approach predictors (KS-Boruta). A corrected paired t-test was utilized to compare six validated models (area-Boruta, tree-Boruta, voxel-Boruta, KS-Boruta, KS-all, ridge-all) for each forest attribute. Based on significant reductions in error (SMdAPE), basal area and biomass were best modeled with KS-Boruta while volume was best modeled with KS-all. Analysis of variable importance shows voxel-based predictors are critical for the prediction of the three forest attributes. This study highlights the importance of multi-resolution voxel-based variables for modeling forest attributes in an area-based context.

Lost in the woods: Forest vegetation, and not topography, most affects the connectivity of mesh radio networks for public safety

“…The percentage of returns classified as “ground” was also calculated using the lidR cloud_metrics function using all vegetation returns ≥ 0.27 m and the lidR LAD function was used to calculate leaf area density using 1 m height bins from 2.5 m to 30.5 m using all, first, and last vegetation returns ≥ 0.27 m ( Table 1 ). Finally, custom functions [ 83 ] developed by Blackburn et al [ 84 ] were used to calculate a variety of cloud- and voxel-based metrics ( Table 1 ). Both the point cloud, or area-based, and voxel-based variables summarized aspects of point density, height, and intensity across each section.…”

“…In total, 919 lidar metrics were calculated. These lidar DEM-, point cloud-, and voxel-derived metrics were included because they have been shown to be correlated with forest structural parameters [ 78 , 84 , 85 , 87 – 103 ].…”

“…Because so many predictor variables were calculated using the remote sensing data, the Boruta algorithm implemented in the R Boruta package (version 7.0.0) [ 136 ] was used to select candidate predictors for each normalized variable set (LIDSAT, LID, SAT). The Boruta algorithm is a feature selection method based on the random forest algorithm [ 136 ] and has been used effectively with high dimensional remote sensing data [ 84 , 137 ]. The Boruta algorithm was run 100 times for each of the six proportion responses (i.e., connectivity levels) and for each of the three variable sets using a maximum number of 1,000 runs each time.…”

“…This function calculates Martone's modified dissection [108,112] DEM Topography PLOS ONE using all, first, and last vegetation returns � 0.27 m (Table 1). Finally, custom functions [83] developed by Blackburn et al [84] were used to calculate a variety of cloud-and voxel-based metrics (Table 1). Both the point cloud, or area-based, and voxel-based variables summarized aspects of point density, height, and intensity across each section.…”

“…Voxels, or volumetric pixels, are created by dividing the lidar point cloud along the horizontal and vertical axes, and voxel- Height-based descriptive statistics within a voxel (s1) and summarized at the section level (s2) for a given resolution (res = 3m, 4m, 5m). Statistics (s1 and s2) = median, mean, variance, standard deviation, coefficient of variation, IQR, skewness, and kurtosis [84,85] Voxel Height i_s1_s2_res Intensity-based descriptive statistics within a voxel (s1) and summarized at the section level (s2) for a given resolution (res = 3m, 4m, 5m). Statistics (s1 and s2) = median, mean, variance, standard deviation, coefficient of variation, IQR, skewness, and kurtosis [84,85] Voxel Intensity…”