A Methodology to Examine the Strengths and Limitations of using FIA Plot Data and LiDAR-based Structural Metrics for Continuous Mapping in the Sierra Nevada Ecoregion

Abstract

Model-based estimation with airborne LiDAR and the US Forest Service’s Forest Inventory and Analysis (FIA) program is a common way to estimate above-ground carbon (AGC), basal area (BA), and wood volume (WV) in forested areas. This study investigates how well the FIA plots capture forested areas within the Sierra Nevada ecoregion and what structures and conditions they fail to capture for effective model-based estimation of AGC, BA, and WV. Random Forest models were used to model these metrics and had a high coefficient of determination (0.79, 0.70, and 0.79 respectively). However, due to strict FIA confidentiality requirements of the true plot locations, the results of this study use the publicly available FIA plot locations, which are guaranteed to be within 1 mile (1.6 km) of the true locations. Results from this study therefore demonstrate the methods, types of results and analysis that could be performed when the true, confidential plot locations are available. Random Forest models with the public location had relatively low coefficients of determination (0.37, 0.34, and 0.36, respectively). The poor performance our models is likely due to the misalignment between remote sensing and plot locations. Using Meyer and Pebesma’s (2021) concept of area of applicability, the area to which a model can be reliably applied without risking overextrapolation, we determined the effectiveness of the FIA sampling scheme for representing structures across the ecoregion for modeling purposes. Based on an evaluation of area of applicability, FIA plots captured 96% of ecoregion for each forest metric, respectively. We found that we tall forests across a range of canopy covers are not well modeled by the FIA plots. These include very tall, very dense forests with dominant tree height greater than 45m and canopy cover percent more than 50% as well as forests with tall trees greater than 30 m where canopy cover is low. Forests that have recently experienced disturbance are also underrepresented. Model-based estimation of AGC, BA, and WV with FIA plots and LiDAR is effective across much of the ecoregion, but there are some systematic under sampled forest types, structures and conditions where our models may be biased. Overall, this study highlights the importance of quantifying the area in which a sample can be effectively applied for model-based estimation. Without proper quantification of where a map may be biased, researchers, forest managers, and policy makers may unknowingly use maps over areas where they may not be appropriate. Results of this study are preliminary and pending the use of the true FIA coordinates. Regardless, this study provides a valuable demonstration of a workflow to characterize how well a sampling scheme represents an area of interest.