Application and evaluation of topographic correction methods to improve land cover mapping using object-based classification

This study applies and evaluates topographic correction methods to reduce radiometric variation due to topography characteristics in rugged terrain. The aim of this study was to improve the capability of satellite images to generate more reliable land cover mapping using object-based classification. Several semi-empirical correction methods, which require the estimation of empirically defined parameters, were selected for this study. Usually, these parameters are estimated relying on a previous land cover map. However, in this work the correction methods were applied considering the unavailability of a previous land cover map and the ease for implementation, so the main land cover type was used to estimate correction parameters to be applied to correct all land cover type. Landsat 5 TM image and topographic data derived from SRTM (Shuttle Radar Topography Mission) over an area located in an agricultural region of southeastern Brazil were used. Land cover classification was carried out using an object-based approach, which includes image segmentation and decision tree classification. The evaluation of topographic correction methods was based on: spectral characteristics expressed by standard deviation and mean values of spectral data within land cover classes; relationship between spectral data and solar illumination angle on the slope (cos i); object (segment) mean size; decision tree structure; visual analysis; and classification accuracy. Results show that the standard deviation of spectral data and the correlation between spectral values and cos i decreased after data correction, but not for all methods for some of the tested TM bands. The methods herein referred as Cosine, S1, Ad2S and SCS methods showed to increase the standard deviation and the correlation compared to the uncorrected data, mainly for bands 1, 2 and 3. Object mean size, in general, decreased after correction, except for C method. The effect on the object size showed to be related to a calculated standard deviation of adjacent pixels values. The decision tree structure given by the number of leaves also decreased after correction. The C, SCS + C and Minnaert methods showed the highest performance, followed by S2 and E-Stat, with a general accuracy increase around 10%. Land cover classification from uncorrected and corrected data differed in a large portion of the total studied area, with values around 29% for all correction methods. (C) 2014 Elsevier B.V. All rights reserved.