Influence of Lossy Compressed DEM on Radiometric Correction for Land Cover Classification of Remote Sensing Images

World coverage Digital Elevation Models (DEM) have progressively increased their spatial resolution (e.g., ETOPO, SRTM, or Aster GDEM) and, consequently, their storage requirements. On the other hand, lossy data compression facilitates accessing, sharing and transmitting large spatial datasets in environments with limited storage. However, since lossy compression modifies the original information, rigorous studies are needed to understand its effects and consequences. The present work analyzes the influence of DEM quality -modified by lossy compression-, on the radiometric correction of remote sensing imagery, and the eventual propagation of the uncertainty in the resulting land cover classification. Radiometric correction is usually composed of two parts: atmospheric correction and topographical correction. For topographical correction, DEM provides the altimetry information that allows modeling the incidence radiation on terrain surface (cast shadows, self shadows, etc). To quantify the effects of the DEM lossy compression on the radiometric correction, we use radiometrically corrected images for classification purposes, and compare the accuracy of two standard coding techniques for a wide range of compression ratios. The DEM has been obtained by resampling the DEM v.2 of Catalonia (ICC), originally having 15 m resolution, to the Landsat TM resolution. The Aster DEM has been used to fill the gaps beyond the administrative limits of Catalonia. The DEM has been lossy compressed with two coding standards at compression ratios 5:1, 10:1, 20:1, 100:1 and 200:1. The employed coding standards have been JPEG2000 and CCSDS-IDC; the former is an international ISO/ITU-T standard for almost any type of images, while the latter is a recommendation of the CCSDS consortium for mono-component remote sensing images. Both techniques are wavelet-based followed by an entropy-coding stage. Also, for large compression ratios, both techniques need a post processing for correctly delimiting coastline, avoiding the confusion between elevation and no-data values. Six (from March 2005 to May 2007) geometrically corrected Landsat-5 images on the path-row 197-031 have been used. The six optical bands and the NDVI for each date have been introduced in a powerful hybrid classification process. The training areas and the ground truth have been obtained from the Mapa de Cobertes del Sòl de Catalunya (v. 3), a land cover map created by photointerpretation of 0.5 m orthophotomaps acquired between 2005 and 2007 and covering all the extension of Catalonia. The legend has been reduced from 233 categories to 21. Preliminary results have shown that the effect on land cover classification of applying lossy compression to the DEM used in the radiometric correction is small (lower than 1%) even for compression ratios up to 200:1. Comparing classification performance after a compression of 5:1 and and a compression of 200:1 with both coding standards showed that: a) the percentage of correctly classified image was 73%; b) 20% was wrongly classified; c) 3.5% was wrongly classified at compression ratio 5:1; and d) also 3.5% was wrongly classified at compression ratio 200:1. These results are the first in the literature to analyze the effect of DEM lossy compressing when DEM are employed for radiometric correction.