Deriving land cover information from multi-temporal MOS data

Abstract

Since the advent of the "Multispectral Scanner" (MSS) on Landsat, which was operated from 1972 to 1987, mono-temporal optical data have been frequently used for land cover inventories. The use of both multi-spectral and multi-temporal characteristics of the remote sensing data for environmental monitoring has only recently received attention due to the shorter revisit time of satellites like TERRA with the sensor MODIS (MODerate Resolution Imaging Spectroradiometer) and in the near future ENVISAT with the sensor MERIS (MEdium Resolution Imaging Specrometer) on board. In preparing for the use of higher level data products derived from MERIS level 1 data, an automated multi-temporal land cover classification (LCC) methodology is being developed. Prior to the launch of MERIS, the LCC algorithm was tested using MOS (Modular Optoelectronic Scanner) data in order to assess its accuracy and precision. These data are comparable with MERIS data due to the spectral similarity of the sensor, their orbits, and the spatial resolution (~500m for MOS and ~300m for MERIS). All MOS data from 1997 (about 300 scenes of 180 x 180 km), with less than 30% cloud cover, were used for testing the automated multi-temporal land cover classification (LCC) methodology. The scenes were atmospherically corrected using the atmospheric correction model (ATCOR) developed at the German Aerospace Center (DLR). After atmospheric correction and georeferencing the data were composited for ten day periods. An automated supervised maximum likelihood classification was performed for each ten day composite. These mono-temporal classification results were merged to a single multi-temporal classification data set using a "voting" procedure. This approach is based on the phenology of the classes. The resulting information was labelled according to the IGBP-DISCover legend. The accuracy of the resulting LCC map was assessed for the German land area using CORINE land cover data as a reference. It was shown that the classification results using the multi-temporal procedure are more reliable, and often exceed the best mono-temporal accuracies and reliabilities. Despite the shortcomings of the MOS instrument, in comparison with the swath width and revisit times of MERIS, MOS data enabled us to verify the potential success of this fairly simple and robust approach towards automated land cover mapping.