Using topographic attributes to predict the 1 understorey structure of a wet eucalypt forest 2

Background: Forest understorey structure is an important component of forest ecosystems that affects forest-dwelling species, nutrient cycling, fire behaviour, 18 biodiversity, and regeneration capacity. Mapping the structure of forest understorey 19 vegetation with field surveys or high-resolution LiDAR data is costly. We tested 20 whether landscape topography and underlying geology could predict the 21 understorey structure of a 19 km 2 area of wet eucalypt primary forest located at the 22 Warra Long Term Ecological Research Supersite, Tasmania, Australia. In this 23 study, we used random forest regressions based on twelve topographic attributes 24 derived from digital terrain models (DTMs) at various resolutions and a geology 25 variable to predict the densities of three understorey layers compared to density 26 estimates from a high resolution (28.66 points/m 2 ) LiDAR survey. 27 Results: We predicted the vegetation density of three canopy strata with a high 28 degree of accuracy (validation root mean square error ranged from 8.97% to 29 13.69%). 30 m resolution DTMs provided greater predictive accuracy than DTMs 30 with higher spatial resolution. Variable importance depended on spatial 31 resolutions and canopy strata layers, but among the predictor variables, geology 32 generally produced the highest predictive importance followed by solar radiation. 33 Topographic position index, aspect, and SAGA wetness index had moderate 34 importance. LiDAR point clouds. This study should help in assessing fuel loads, carbon stores, 40 biomass, and biological diversity, and could be useful for foresters and ecologists 41 contributing to the planning of sustainable forest management and biodiversity 42 conservation. 43

understoreys is a vital attribute that affects habitat quality, including foraging and 53 breeding resources, for many forest animals (Camprodon and Brotons 2006). 54 Thus, characteristics of vertical structure could be used to develop quantitative 55 6 A 19.09 km 2 core area of mature native forest was used for analysis after 109 excluding 2.79 km 2 of roads, rivers, and previously harvested sites and 3.12 km 2 110 of edges (Fig. 1). We utilised airborne LiDAR and a geology data layer. High-111 resolution LiDAR data was collected with a Riegl LMS-Q560 sensor scanning at 8 response variables since the high-resolution LiDAR data will produce a 'best case 138 scenario' for quantifying relationships between topography and forest structure 139 and enables us to compare DTM resolutions without introducing other biases. 140 Assessing the predictive capacity of other freely available DTMs was beyond the 141 scope of this study. The topography and geology datasets provided thirteen 142 predictor variables, which were deployed for random forest regression modelling 143 to determine whether topography and geology predict local-scale forest structure. 144

Analysis of RF parameters and variable importance
Firstly, LiDAR point clouds were classified into the ground and non-ground 150 classes using lasground, then normalized using lasheight. Outlier points (above 151 the canopy and below ground level) were removed. 152 The structure of understorey layers was characterised using lascanopy for five 153 different spatial resolutions (1 m, 5 m, 10 m, 20 m, and 30 m). The structure of 154 each layer was characterised as counts of points falling into the height intervals 155 divided by the total number of points within each grid cell. Field data from sample 156 plots and field experience indicate that most emergent overstorey trees were above 157 ~50 m in height, so the three understorey layers were classified as a lower layer (≥ 158 2 to ≤ 10 m), middle layer (> 10 to ≤ 30 m), and an upper layer (> 30 to ≤ 50 m). 159 The canopy at the Warra silviculture system trials (far north-east of landscape)  Calculated by fitting a plane to the eight neighbouring cells (Travis et al. 1975).

Aspect
The orientation of the cell relative to the north (Travis et al. 1975).

Catchment area
The upstream area of each cell (Kiss 2004).
Profile curvature The rate of change of slope in a downslope direction: a proxy for acceleration and deceleration of water over the terrain (Wilson et al. 2007).
Plan curvature The curvature of a contour at the central pixel. It can be used as a proxy for convergence and divergence of water (Wilson et al. 2007).

LS (slope length and steepness) factor
Calculated using the upslope catchment area of each cell and the grid cell slope. This metric is used as a proxy for erosivity (Desmet and Govers 1996;Kinnell 2005).

Potential solar radiation ratio
The ratio of the potential solar radiation on a sloping surface to that on a horizontal surface (Moore et al. 1991).

Topographic Position Index
Whether any particular pixel forms part of a positive (e.g., crest) or negative (e.g., trough) feature of the surrounding terrain (Wilson et al. 2007).

Terrain Ruggedness Index
The sum change in elevation between a grid cell and its eight neighbouring grid cells (Riley et al. 1999).

Random forest model prediction and cross-validation 225
For all three canopy layers, models with the coarsest spatial resolution (30 m pixel 226 size) outperformed those at smaller spatial resolutions (1 m, 5 m, 10 m, and 20 m) 227 for both prediction and validation (Table 2). We were unable to explain the 228 pattern of increased error for 10 m pixel size in the lower vegetation layer. The 229 best model accuracy from the validation dataset was for the lower layer and the 230 lowest accuracy for the middle vegetation layer. This contrasts with the prediction 231 dataset, with the best accuracy for the upper layer, although there was little 232 difference in the magnitude with RMSE ranging from 7.61% (R 2 = 0.82) for the 233 upper layer through to 9.76% (R 2 = 0.77) for the middle layer. Hereafter we 234 present results only for the 30 m resolution datasets. 235 LiDAR dataset (Fig. 4). 241 on Jurassic dolerite geology class had dense lower and middle layers and the most 248 sparse upper layer. The Jurassic dolerite was confined to a small area of high 249 elevation within the study site (Fig. 1). By contrast, the Permian sediments 250 geology class had a dense upper layer, and the least dense lower layer (Fig. 6). north-western United States, and they argued that shrub-herb cover is more 315 heterogeneous than overstorey cover attributes. 316

Optimal spatial resolution 317
Our use of a single high-resolution LiDAR dataset to derive a range of spatial 318 resolutions for DTMs allowed us to determine that the optimal spatial resolution 319 for predicting understorey density of different vegetation layers in our study 320