Developing an Ecoregion-level Imputation Model From LiDAR-derived Biomass Maps

Andrew Thomas Hudak, USDA Forest Service, ahudak@fs.fed.us (Presenter)
Patrick A Fekety, University of Minnesota, pafekety@umn.edu
Michael J Falkowski, University of Minnesota, mfalkows@umn.edu
Robert E Kennedy, Oregon State University, rkennedy@coas.oregonstate.edu
Alistair Matthew Stuart Smith, University of Idaho, alistair@uidaho.edu

Accurate maps of aboveground biomass (AGB) are valuable tools for estimating carbon storage and large scale planning. LiDAR-based prediction models are useful for generating accurate maps of desired attributes, such as biomass, but their usefulness is limited to areas where LiDAR have been collected. Our objective is to upscale landscape-level, lidar-derived AGB maps to the regional level. Previous work used 285 field plots and four LiDAR collections from northern Idaho to create a LiDAR-derived AGB model (root mean square distance, RMSD=90 Mg/ha, (42%)) that was applicable to the Northern Rocky Mountain Province of the Bailey Ecoregion System. In this study, LiDAR-derived AGB maps from the previous work were used as training data for Most Similar Neighbor (MSN) imputation models. Predictor variables included Landsat-derived brightness, greenness, and wetness tasseled cap values, climate metrics, and topographic metrics. The MSN models used values from ~800,000 pixels as reference observations and predicted AGB at a 30-meter spatial resolution across the four LiDAR extents. The effect of the number of nearest neighbors on the imputation results was investigated. Preliminary results showed that increasing the number of nearest neighbors reduced the RMSD and improved the fit statistics for the line of best fit between observed and predicted AGB (k=1: RMSD=137 Mg/ha (74%), b0=100 (i.e. intercept), b1=0.46 (i.e. slope), R2=0.22; k=10: RMSD=102 Mg/ha (55%), b0=19, b1=0.90, R2=0.42; k=100: RMSD=98 Mg/ha (53%), b0=2, b1=0.99, R2=0.46). These results suggest it will be possible to apply a MSN model across the Northern Rocky Mountain Province and generate a fine scale AGB map. The disadvantage with increasing k is that the variance of predictions becomes reduced relative to observations. Further work is needed to identify the ideal number of nearest neighbors that should be used, and to what extent the mapped regional AGB predictions from the various models may be biased when averaged within county-level polygons, as validated with Forest Inventory and Analysis (FIA) summary data.

Presentation Type:  Poster

Session:  Carbon Monitoring System (CMS) Posters   (Mon 1:30 PM)

Associated Project(s): 

• Hudak, Andy: Prototyping A Methodology To Develop Regional-Scale Forest Aboveground Biomass Carbon Maps Predicted From Landsat Time Series, Trained From Field and Lidar Data Collections, And Independently Validated With FIA Data ...details

Poster Location ID: 155