A New Approach to Categorical Resampling

A new method has been developed for resampling raster image data that contain class or categorical data. Categorical data are usually the result of an image classification or other statistical processes. During reprojection and resampling, the combination or interpolation of data with their neighboring pixels is not necessarily meaningful as it is with signalbased remote sensing data. The nearest neighbor resampling method is commonly used to resample this type of data. This method is chosen because the alternatives--cubic convolution, bilinear interpolation, etc. -are interpolating methods that do not preserve categorical values. In addition, the geometric distortions that are present in the projection change of data of global extent are far greater than distortions that occur in moderateto high-resolution remote sensing data. Indeed, many of the software tools available today were designed for single-scene, signal-based remote sensing image data where the extent of the image is usually only a few hundred kilometers, rather than for datasets of global or continental coverage. The typical nearest neighbor resampling algorithm for categorical data takes into account only the center of the pixel and not the area covered by the pixel. In instances where the image (or a region of the image) is being undersampled, nearest neighbor resampling can result in imagery that is not representative of the original image. The new resampling method treats pixels not as points, but as areas. This approach maps the corner coordinates of the output image pixel back into the input image and statistically determines the pixel value on the basis of input image pixels that lie within the output pixel's geometric extent. 1 USGS EROS Data Center, SAIC, Sioux Falls, SD 57198-0001. Work performed under U.S. Geological Survey contract 03CRCN0001.