Unmixing-based Landsat and Meris Image Fusion for Land Cover Mapping over the Netherlands

Anthropogenic land use activities are significantly contributing to the ecological degradation of the Earth system. Therefore, having actual and reliable land cover information is fundamental to study the impact of such an ecological degradation on our future welfare. High spatial resolution sensors, such as Landsat TM, are typically used to derive land cover information from local to regional scales. However, current high spatial resolution sensors do not provide an appropriate temporal resolution. This is especially true for areas having high cloud coverage throughout all the year. In this respect, The Medium Resolution Imaging Spectrometer, MERIS, aboard the ESA-Envisat environmental satellite delivers data every 2-3 days. This increases the chances of encountering cloud free regions. Nevertheless, MERIS works at a spatial resolution of 300m (full resolution mode), which might not be sufficient to capture the details of highly fragmented landscapes. This is why the synergic use of these 2 sensors was explored in this paper. An unmixing-based data fusion technique was used to generate images with the spatial resolution of Landsat TM and with the spectral (and eventually temporal) resolution provided by MERIS. More precisely, one Landsat TM and one MERIS full resolution image acquired in July 2003 over The Netherlands were fused using the linear spectral mixing model. First an unsupervised classification of the Landsat TM image was done to obtain the fractional coverages of the different land cover types present in each MERIS pixel. Next, the spectral signatures of each land cover type were retrieved by “inverting” the linear mixing model. This is, MERIS “endmembers” were obtained from the known fractional coverages of each pixel. After that, both the fused and the Landsat TM images were classified to produce maps of the 8 main land cover types over The Netherlands. These maps were subsequently validated using the Dutch land use spatial database (LGN5) as a reference. The paper concludes by describing the potentials and limitations of this multi-sensor approach with respect to the solely use of Landsat TM data. 1 Corresponding author: [email protected]