Page 17 - Interp Book
P. 17

Remote Sensing Classification

               We used three dates of Landsat Thematic Mapper satellite data, combined with other
               information, to classify the landcover.  After data acquisition, the next step in the
               classification process was to create a seamless mosaic of LandSat scenes for all dates.

               The generation of the mosaics was neither a straightforward nor a simple task.  The
               imagery used to build the mosaics needed to be, for the most part, cloud-free.  This
               condition rarely exists in practice.  Because of clouds, often a given path-row of imagery
               was itself a mosaic.  Maintaining a consistent date throughout each seasonal mosaic
               brought additional complexity to the process.  The most challenging step in generating
               the seasonal mosaics was the issue of color balancing.  This process removes the
               apparent divisions among adjacent path-rows of imagery by matching, on a band by
               band basis, the histograms of all the images used.  This process is iterative in nature
               and is often one of the most labor intensive portions of the landcover mapping protocol.

               We used a decision tree classification approach to classify the initial 14 landcover
               classes for Phase 1 (Table 2).  In Phase 2, Pine Plantation was removed as a possible
               target for the classification, but Shinnery Oak was added, resulting in 14 landcover
               classes for Phase 2 as well. This approach allows for the combination of remotely
               sensed data with ancillary data in a flexible way.  We tried multiple different
               combinations of satellite reflectance data and ancillary data before settling on a final
               combination that provided the best result.  Important ancillary data used for
               classification (in addition to all 6 LandSat reflectance bands for three dates), included
               slope, aspect, landscape position, solar insolation, percent canopy cover from the
               National Landcover Dataset (NLCD), percent impervious surface from the NLCD,
               change detection from the NLCD, and agricultural areas as defined by the most recent
               version of the National Agricultural Statistics Service cropland data layer.


























                                                              9
   12   13   14   15   16   17   18   19   20   21   22