Estimating stand age from airborne laser scanning data to improve ecosite-based models of black spruce wood quality in the Boreal forest of Ontario
Rebecca Wylie & Jeff Dech – Nipissing University: Topic 2, Question 14
Models that provide reliable estimates of wood quality from source data available in forest inventories could enable value chain optimization approaches that consider the market potential of trees prior to harvest. Ecological land classification units (e.g. ecosite) and forest structural metrics derived from Light Detection and Ranging (LiDAR) data have been shown to be useful predictors of wood quality, which capture variations related to site quality and forest development. Such indicators can be used to classify stands according to the predicted attributes of average trees, and identify portions of the landscape that are likely to contain trees with a desired suite of characteristics. However, much of the variation in wood quality among trees is driven by differences in age, and this variation remains unaccounted for in models because age is poorly represented in most inventory systems. This study was conducted in the Hearst Forest in northeastern Ontario, across a network of 166 plots representing a broad range of site conditions and forest structure. Plots were also linked to a raster (20 x 20m) of LiDAR derived structural variables covering the entire forest. Increment cores collected from these sites to attain age and wood quality information were from representative, dominant and codominant black spruce trees.
Modeling Black Spruce Stem Age:
Our first objective was to use LiDAR-derived forest structure and site metrics to predict mean stem age of black spruce from a subset of 116 plots which best represented black spruce-dominated stands. Modelling age in the Hearst forest is challenging due to the large range of mean stand ages in our data set (11-160 years) and large standard deviation (1-69 years) of age within some plots. Particularly, plots with even small amounts of hardwood species, or cedar in combination with lower dominance of black spruce (e.g. < 70%), have large prediction errors. After employing a variety of parametric and nonparametric modeling techniques, our best stem age model was produced using an approach that combined the k nearest neighbour (KNN) approach with random forests as the distance metric. This model has a root mean squared distance of 16.7 years and explains 56% of the variation in the sample population. Overall, the model used LiDAR structural metrics describing height, canopy closure and site quality to gain a more representative estimate of mean stand age than traditional FRI methods.
Modeling Black Spruce Stem Density, MOE and MFA:
Our second objective was to improve models of black spruce wood quality by including age, as a predictor variable. Three key wood qualities, basic density, modulus of elasticity (MOE), and microfibril angle (MFA), were the responses of interest. We used data from 169 increment cores and ran the analysis with regression trees and Random forests simulations. Variables were chosen using stepwise multiple linear regression and variation inflation factor (VIF) to insure the model was not overfit. For all three predictor variables, age was consistently one of the most important variables. For example, the first split in regression tree of wood density separated samples with plot ages less than 35 years, showing the lowest wood density appearing in the youngest trees (Figure 1). As expected ecosite also had a large influence on wood quality, explaining high density of tree fibers with small dbh in conifer swamps, EG8r, EG8i and EG8p. Overall the regression tree explained 46% of the variation of wood density at breast height, 47% of the variation of MFA and 54% of the variation in MOE in the sample population. Random forests returned similar results showing that age reduced mean square error (mse) by 60% and ecosite group reduced mse by 50%.
Figure 1. Regression tree analysis for 169 wood quality samples describing density, using plot age, quadratic mean diameter, ecosite group (a= EG2-Dry sandy, b= EG3Fresh sandy or dry to fresh coarse loamy, c=EG4 moist sandy to coarse loamy, d=EG5 fresh clayey, e=EG6 fresh silty to fine loamy, f=EG7 moist silty to fine loamy to clayey, g= EG8r rich conifer swamps, h=EG8i intermediate conifer swamps and h= EG8p poor conifer swamps), as well as airborne LiDAR derived metrics representing canopy closure at 2m and 12 m (cc2, cc12), and cumulative % of LiDAR returns between 2-4m above the ground(D2). At each branch in the tree the classification moves left if the condition is met.
The next phase of this project will use the imputed ages from our KNN model and the subsequent classification derived from our random forests analysis to map mean wood qualities across representative stands in the Hearst forest. These maps could be used to plan harvest according to market demands, and to limit waste. Another benefit could be adding mean stand level density to carbon models to better account for the effect of site conditions on carbon storage. Although our new models do not show improved prediction accuracy of average tree properties with the addition of age, the expanded models increase the range of forest conditions over which black spruce wood quality can be predicted.