LiDAR based prediction of forest biomass using hierarchical models with spatially varying coefficients

REMOTE SENSING OF ENVIRONMENT

卷:169

Article

Babcock, Chad¹  Finley, Andrew O.^2,3  Bradford, John B.⁴  Kolka, Randall⁵  Birdsey, Richard⁶  Ryan, Michael G.^7,8 

[1] Univ Washington, Sch Environm & Forest Sci, Seattle, WA 98195 USA

[2] Michigan State Univ, Dept Forestry, E Lansing, MI 48824 USA

[3] Michigan State Univ, Dept Geog, E Lansing, MI 48824 USA

[4] US Geol Survey, Southwest Biol Sci Ctr, Flagstaff, AZ 86001 USA

[5] US Forest Serv, USDA, No Res Stn, Grand Rapids, MN USA

[6] US Forest Serv, USDA, No Res Stn, Newtown Sq, PA USA

[7] Colorado State Univ, Nat Resources Ecol Lab, Ft Collins, CO 80523 USA

[8] US Forest Serv, USDA, Rocky Mt Res Stn, Ft Collins, CO USA

关键词: Bayesian hierarchical models;    Markov chain Monte Carlo;    Gaussian process;    Geospatial;    LiDAR;    Forest biomass;   

DOI  :  10.1016/j.rse.2015.07.028

来源: Elsevier

【 摘 要 】

Many studies and production inventory systems have shown the utility of coupling covariates derived from Light Detection and Ranging (LiDAR) data with forest variables measured on georeferenced inventory plots through regression models. The objective of this study was to propose and assess the use of a Bayesian hierarchical modeling framework that accommodates both residual spatial dependence and non-stationarity of model covariates through the introduction of spatial random effects. We explored this objective using four forest inventory datasets that are part of the North American Carbon Program, each comprising point-referenced measures of above-ground forest biomass and discrete LiDAR. For each dataset, we considered at least five regression model specifications of varying complexity. Models were assessed based on goodness of fit criteria and predictive performance using a 10-fold cross-validation procedure. Results showed that the addition of spatial random effects to the regression model intercept improved fit and predictive performance in the presence of substantial residual spatial dependence. Additionally, in some cases, allowing either some or all regression slope parameters to vary spatially, via the addition of spatial random effects, further improved model fit and predictive performance. In other instances, models showed improved fit but decreased predictive performance indicating over-fitting and underscoring the need for cross-validation to assess predictive ability. The proposed Bayesian modeling framework provided access to pixel-level posterior predictive distributions that were useful for uncertainty mapping, diagnosing spatial extrapolation issues, revealing missing model covariates, and discovering locally significant parameters. (C) 2015 Elsevier Inc. All rights reserved.

【 授权许可】

Free