Spatial structure and landscape associations of SRTM error

Abstract

This paper evaluates the spatial structure of Shuttle Radar Topography Mission (SRTM) error and its associations with globally available topographic and land cover variables across a wide range of landscapes. Two continental-scale SRTM elevation data samples were extracted, along with collocated National Elevation Dataset (NED) elevations, MODIS composite forest cover percentage, and global ecoregion major habitat type codes. The larger punctual sample contained nearly 247,000 sites on a regular grid across the conterminous United States, while the smaller areal sample consisted of 37,500 45″ × 45″ rectangular regions on a regular grid. Sub-pixel positional mismatch was accounted for by finding and using the best local fit between the 1 arc sec horizontal resolution NED product and the 3 arc sec (3″) horizontal resolution SRTM product. Slope and aspect were calculated for all samples. Using the larger point sample, we identified associations between SRTM error, defined as NED–SRTM 3″ differences, with these land cover and terrain derivative variables. Using the areal sample, we developed semivariograms of elevation error for tens of thousands of small regions across the United States, as well as for sets of these regions with common slope and landcover properties. This facilitated a more comprehensive evaluation of the spatial structure of SRTM error than has previously been done. The punctual sample RMSE was 8.6 m, conforming to previous estimates of SRTM error, but many errors in excess of 50 m were identified. Nearly 90% of these large errors were positive and correlated with high forest cover percentage. Overall, SRTM elevations consistently overestimated the surface. Forest cover and slope were positively correlated with positive bias. A strong association of aspect with SRTM error was noted, with positive error magnitudes peaking for aspects oriented to the northwest and negative error magnitudes peaking for slopes facing southeast. Error bias, standard deviation, and semivariograms differed substantially across ecoregion types. These variables were incorporated in a regression model to predict SRTM error: this model explained nearly 60% of the total error variation and has the potential to substantially improve the SRTM data product worldwide using globally available datasets.