A Hierarchical Model for Quantifying Forest Variables Over Large Heterogeneous Landscapes With Uncertain Forest Areas

Area of Science:

• Forestry
• Spatial Statistics
• Remote Sensing

Background:

• Predicting continuous forest variables (e.g., biomass, volume) at fine resolutions is crucial.
• Current two-step prediction methods often ignore errors in forest/non-forest classification, leading to biased estimates.
• Uncertainty propagation from forest status prediction to variable prediction is rarely addressed.

Purpose of the Study:

• Develop a modeling framework to propagate uncertainty in forest/non-forest prediction.
• Improve the accuracy of continuous forest variable predictions by accounting for classification errors.
• Address the challenge of biased estimates in forest resource assessments.

Main Methods:

• Proposed a novel modeling framework with two latent processes: a continuous spatial process for forest variables and a binary spatial process for forest presence.
• Applied georeferenced National Forest Inventory (NFI) data and remote sensing predictors.
• Utilized a low-rank predictive process for dimension reduction to handle large datasets and computational burden.

Main Results:

• The proposed framework allows for the propagation of uncertainty from forest status prediction.
• Demonstrated the adaptation of low-rank predictive processes for efficient modeling of large spatial datasets.
• Reduced computational burden and dimensionality in spatial forest variable prediction.

Conclusions:

• The developed modeling framework effectively incorporates uncertainty from forest detection, leading to more reliable forest variable predictions.
• Dimension reduction techniques like low-rank predictive processes are vital for computationally intensive spatial analyses.
• This approach enhances the accuracy of forest resource assessments by acknowledging and propagating prediction errors.