ESTIMATING FOREST STRUCTURE AND SURFACE FUEL LOADS IN DRY CONIFER FORESTS OF THE COLORADO FRONT RANGE WITH TERRESTRIAL LASER SCANNING

Abstract

Measurements of forest structure and surface fuel parameters are a fundamental component of forest inventories in the Southwestern United States. Traditionally, these parameters are estimated using manual field-based methods that can be time-consuming and resource-intensive, thereby reducing the sampling intensity and accuracy of inventory campaigns. Recently, active remote sensing approaches have emerged as an alternative to field-based inventory methods. Terrestrial laser scanning (TLS) is one such approach that offers the potential for an increased level of detail and accuracy in inventory data while requiring fewer resources than traditional methods. Previous studies have demonstrated that TLS data can be used to directly measure forest structure parameters and predict surface fuel loads using statistical models. However, to date, no studies have evaluated the accuracy of TLS approaches in the dry conifer forests of the Colorado Front Range. In Chapter 1 of this thesis, I investigate the accuracy of single- and multi-scan TLS estimates of five forest structure parameters (mean diameter at breast height (DBH), mean tree height, canopy base height (CBH), tree density, and basal area) measured at 74 dry conifer-dominated plots within the Colorado Front Range. At each plot, I collected four TLS scans, then used single-scan (one scan) and multi-scan (four scans) data in an automated individual tree segmentation algorithm to estimate structural parameters. To assess accuracy and bias, these estimates were compared to measurements obtained from traditional field-based methods. My results indicate that TLS approaches can accurately estimate forest structure, but accuracy levels differ among parameters. In addition, I found that accuracy differences between scanning approaches never exceeded 2.3% for any parameter, suggesting that the more complex multi-scan approach is unnecessary in dry conifer forests of the Colorado Front Range. Median errors from the single-scan approach were smallest for DBH and height (10.8%-12.3%) and largest for CBH, tree density, and basal area (32.2%-38.7%). Bias estimates showed that DBH, tree density, and basal area were systematically overestimated, whereas height and CBH were underestimated. Additionally, errors were greatest in plots that had both low tree densities and understory occlusion from high shrub/sapling abundance and low crown base heights. These findings suggest that the single-scan approach can provide an accurate and parsimonious inventory method in forests with high crown base heights and open understories and may enable greater sampling intensities for lower resource inputs than other approaches. Chapter 2 investigates the predictive accuracy of fine-scale surface fuel models built from TLS metrics and destructively sampled surface fuel load measurements from 128 0.25 m2 volumetric clip-plots. At each clip-plot, I took a single TLS scan, then collected destructive estimates of fuel load for the following five surface fuel components: fine dead surface fuels, large-diameter down dead woody fuels, herbaceous fuels, live woody fuels, and total surface fuels. Three predictive models (random forest, elastic net, and multiple linear regression via leaps and bounds selection) were developed for each component using 166 TLS point cloud-derived predictor variables and a 10-repeat 5-fold cross-validation scheme. My results showed that predictions were relatively unbiased; however, normalized root mean square error (NRMSE; 60%-548%) and explanatory power (R2 = -11.16-0.37) were variable across modeling approaches, fuel components, and cross‑validation runs. In general, the elastic net model produced the lowest errors, followed by random forest, then leaps and bounds models. Errors across modeling approaches were smallest and most consistent for fine dead surface fuels (NRMSE = 60%-62%), but large and highly variable for large-diameter down dead wood (NRMSE = 315%-548%). These findings indicate that TLS-based models can be error-prone and inconsistent when predicting fine-scale fuel loading in Colorado Front Range dry conifer forests and highlight the need for additional research to improve accuracy.