Incorporating Remote Sensing into Forest Carbon Accounting using Model-based and Model-assisted Estimators

Abstract

Forests’ ability to sequester carbon dioxide (CO2) from the atmosphere has been iden-tified as a solution to mitigate anthropogenically caused climate change. The impor- tant role forests play in storing carbon has elevated the need to quantify aboveground biomass (AGB) stocks. Conventional estimation of AGB often uses probabilistically sampled field data to make population estimates of AGB. Estimates with the desired level of precision from field-only inventories typically require an intensive sampling de- sign that is expensive to collect. A method to increase estimation precision or reduce the number of needed field sample plots is to incorporate remote sensing information into AGB inventories. Using both the design-based and model-based inferential paradigm, the relationship between observed AGB and remote sensing variables can be exploited to increase estimation precision compared to traditional field-only inventories. This thesis details two projects that leverage remote sensing data to improve estimation precision for AGB as well as other forest inventory variables. Project 1: Crowd-sourcing field and remote sensing datasets is an excellent way to accumulate the training information needed to create regional AGB maps using machine learning (ML). However, ML-based estimators can be biased when they are trained using remote sensing data collected with multiple different sensor systems and field data collected with multiple different sampling protocols. Here, we demonstrate how to correct for potential ML estimator bias using model-assisted (MA) and geostatistical- model-based (GMB) estimators. Using probabilistically sampled field AGB values from the USFS Forest Inventory and Analysis (FIA) Program as the response variable and ML predictions as the predictor variable, we show how MA and GMB estimators can correct for ML estimator bias and be used to generate statistically defensible uncertainty estimates. Our motivating dataset is a collection of AGB maps generated using Random Forest (RF-AGB) and a probability sample of AGB values for Oregon from 2001 to 2016. Using the maps generated from the RF-AGB estimator, we apply MA and GMB estimators to generate areal density estimates at the state and county level. We also explore a case study at the HJ Andrews Experimental Forest using the GMB estimator at smaller scales than the county level. Results show that the proposed AGB density estimators that leverage the RF-AGB predictions are more precise those based on field data alone. Project 2: Inclusion of remote sensing data into forest inventories has been pro- posed to help improve estimation precision of AGB and other forest inventory variables, a method known as an enhanced forest inventory. An under-represented remote sens- ing technology in enhanced forest inventories is terrestrial laser scanning (TLS). TLS data are collected from the ground, making it challenging to collect wall-to-wall for an entire forest. Sampling strategies can be used to collect TLS data for a representative subset of locations within the forest. In this study we make use of a model-assisted double sampling estimator to relate TLS data to field measured AGB, basal area and tree density. We also include a canopy height model derived using digital aerial pho- togrammetry and the normalized difference vegetation index to assess improvements in estimation precision compared to using field data alone.