Estimating forest area using national forest inventory data, percent forest cover maps, and zero-one-inflated beta regression

Abstract

Accurate and consistent forest area information is critical for environmental monitoring, resource management, and policy development. However, common methods for mapping forest area often rely on binary classification masks and inconsistent definitions, limiting their interpretability and accuracy. Remote sensing products frequently conflict with field-based inventories, and the uncertainty introduced by these differences is rarely quantified. This thesis addresses these challenges by proposing a modeling framework that directly estimates forest area proportion while incorporating uncertainty. A zero-one-inflated beta regression model (ZOIB) was constructed using Forest Inventory Analysis (FIA) field data and National Land Cover Database tree canopy cover remote sensing data through a hierarchical Bayesian framework. The FIA forest area proportion data are heavily zero-one-inflated, with a large proportion of values exactly at 0 or 1, corresponding to fully forested or fully non-forested area. Because of this unusual distribution of forest area, it is important to use a model that can accurately capture its shape and characteristics. The ZOIB model was compared with a Beta regression model with no zero-one inflation and random forests regression model (RF). For the ZOIB and Beta models, posterior predictive distribution summary statistics of mean and median were compared, along with the mean and median of the RF model outputs. The Beta model performed worst among the models, and mean performed poorly against the median, both primarily due to their inability to reflect zero-one-inflated data. The median RF model was the most accurate, with a mean absolute error of 0.1530, followed by the median ZOIB model at 0.1630. Given its balance of predictive performance and interpretability, the median ZOIB model was selected as the final modeling approach. While slightly less accurate, the ZOIB model offered the advantage of consistent, interpretable predictions along with full uncertainty quantification. The ZOIB approach presented in this study is particularly well-suited for integration into propagative models such as biomass or carbon estimation, and for supporting more rigorous, data-informed forest management decisions.