The objective of this research is to develop an approach to correct nonlinear errors in the SRTM (Shuttle Radar Topography
Mission) elevations, which cannot be handled by most traditional methods. First, a set of uncorrelated feature attributes has been
generated from the SRTM digital elevation model (DEM) together with the new freely available Sentinel-2 multispectral
imagery, over a dense urban area in Egypt. Second, the SRTM DEM, Sentinel-2 image, and the generated attributes have been
applied as input data in an artificial neural network (ANN) classification model to assign each pixel to each of 12 reference
elevations. Finally, the posterior probabilities obtained for ANN have been combined based on an inverse probability weighted
interpolation (IPWI) approach to estimate revised SRTM elevations. The results were compared with a reference DEMwith 1-m
vertical accuracy derived through image matching of the Worldview-1 stereo satellite imagery. The process of performance
evaluation is based on various statistics such as scatter plots, correlation coefficient (R), standard deviation (SD), and root mean
square error (RMSE). The results show that, using the SRTMDEMas a single data source, the RMSE of estimated elevations has
improved to 3.04 m. On the other hand, including the Sentinel-2 image has improved the RMSE of elevations to 2.93 m.
Including the generated attributes as well has improved the estimated RMSE of the elevations to 2.07 m. Compared with the
results from the commonly used multiple linear regression (MLR) method, the improvement in RMSE of the estimated elevations
can reach 45%. |
