2020 ESA Annual Meeting (August 3 - 6)

COS 53 Abstract - Mapping interior Alaska tree canopy cover using Goddard's LiDAR, hyperspectral, and thermal imager

Bonnie Ruefenacht1, Wendy Goetz1, Vicky C. Johnson1, Sean Patterson1, Stacie Bender2, Kevin A. Megown2, Hans-Erik Andersen3 and Bruce Cook4, (1)RedCastle Resources, Onsite contractor to the USDA Forest Service Geospatial Technology and Applications Center, Salt Lake City, UT, (2)USDA Forest Service Geospatial Technology and Applications Center, Salt Lake City, UT, (3)USDA Forest Service, Pacific Northwest Research Station, Seattle, WA, (4)NASA Goddard Space Flight Center
Background/Question/Methods

Traditional methods for mapping tree canopy cover involve building models using field data or photo-interpreted data along with a collection of predictor datasets. Due to limited transportation infrastructure and inclement weather in interior Alaska, conducting field work is extremely difficult and expensive. Suitable aerial photography for photo-interpretation is also lacking. The majority of Alaska’s interior forests, comprising about 15% of all U.S. forest land, have not been inventoried. This study investigated the effectiveness of using G-LiHT data for the mapping of percent tree canopy cover for interior Alaska.

In 2014, 43 G-LiHT flight lines were flown in the Tanana Valley of Alaska. G-LiHT canopy height models with spatial resolutions of 1m were used. All canopy heights above 2m were treated as tree canopy cover. Canopy height data were aggregated and converted to percent tree canopy cover with spatial resolution of 30m resulting in a database of over 6 million percent tree canopy cover pixels. These data were joined to predictor datasets including Sentinel-2, Landsat OLI, elevation, landcover, climate, soil, and various Sentinel-2 and Landsat OLI indices at 30m spatial resolutions. The database was randomly sampled creating modeling and accuracy datasets. Tree canopy cover modeling was done with randomForest.

Results/Conclusions

The randomForest out-of-bag percent variance explained was 84.9% and the root mean squared error was 12.3. Applying the holdout data to the randomForest model gave an average difference between predicted and observed of 7.8 with a standard deviation of 8.6, a root mean squared error of 11.6, and an r2 of 0.86. The majority of the variables deemed to be most important by the randomForest model were the Sentinel-2 bands and Sentinel-2 indices. These model performance metrics are very similar to tree canopy cover modeling metrics obtained for modeling tree canopy cover using more traditional methods in the conterminous U.S. More investigations are needed to improve model performance, but using G-LiHT data does appear to be an effective method to map tree canopy cover especially in difficult to access regions such as interior Alaska.