Estimating tree canopy cover using harmonic regression coefficients derived from multitemporal Landsat data
| Authors: | Jill M. Derwin, Valerie A. Thomas, Randolph H. Wynne, John W. Coulston, Greg C. Liknes, Stacie Bender, Christine E. Blinn, Evan B. Brooks, Bonnie Ruefenacht, Robert Benton, Mark V Finco, Kevin Megown |
| Year: | 2020 |
| Type: | Scientific Journal |
| Station: | Southern Research Station |
| DOI: | https://doi.org/10.1016/j.jag.2019.101985 |
| Source: | International Journal of Applied Earth Observation and Geoinformation |
Abstract
The goal of this study was to evaluate whether harmonic regression coefficients derived using all available cloudfreeobservations in a given Landsat pixel for a three-year period can be used to estimate tree canopy cover
(TCC), and whether models developed using harmonic regression coefficients as predictor variables are better
than models developed using median composite predictor variables, the previous operational standard for the
National Land Cover Database (NLCD). The two study areas in the conterminous USA were as follows: West
(Oregon), bounded by Landsat Worldwide Reference System 2 (WRS-2) paths/rows 43/30, 44/30, and 45/30;
and South (Georgia/South Carolina), bounded by WRS-2 paths/rows 16/37, 17/37, and 18/37. Plot-specific tree
canopy cover (the response variable) was collected by experienced interpreters using a dot grid overlaid on 1m
spatial resolution National Agricultural Imagery Program (NAIP) images at two different times per region, circa
2010 and circa 2014. Random forest model comparisons (using 500 independent model runs for each comparison)
revealed the following (1) harmonic regression coefficients (one harmonic) are better predictors for
every time/region of TCC than median composite focal means and standard deviations (across times/regions,
mean increase in pseudo R2 of 6.7% and mean decrease in RMSE of 1.7% TCC) and (2) harmonic regression
coefficients (one harmonic, from NDVI, SWIR1, and SWIR2), when added to the full suite of median composite
and terrain variables used for the NLCD 2011 product, improve the quality of TCC models for every time/region
(mean increase in pseudo R2 of 3.6% and mean decrease in RMSE of 1.0% TCC). The harmonic regression NDVI
constant was always one of the top four most important predictors across times/regions, and is more correlated
with TCC than the NDVI median composite focal mean. Eigen analysis revealed that there is little to no additional
information in the full suite of predictor variables (47 bands) when compared to the harmonic regression
coefficients alone (using NDVI, SWIR1, and SWIR2; 9 bands), a finding echoed by both model fit statistics and
the resulting maps. We conclude that harmonic regression coefficients derived from Landsat (or, by extension,
other comparable earth resource satellite data) can be used to map TCC, either alone or in combination with
other TCC-related variables.