Automated Processing of Terrestrial Satellite Imagery Using Bayesian Methods

摘要

The typical approach to developing land cover/land use maps with satellite imagery involves defining spectral classes by clustering the image data and making assignment of classes to pixels. Autoclass, a general-purpose Bayesian classifier, was tested in terms of integration into the land cover map-making procedure and improved information content compared to products using other classifiers. Landsat Thematic Mapper (TM) images, ancillary data for interpreting and evaluating Autoclass results, were provided by the United States Geological Survey (USGS) offices in Menlo Park, California and the USGS in Reston, Virginia. The current version of Autoclass is the product of years of development and refinement. It has been successfully used in several application areas from astronomy to biology. The LISP version of AutoClass was tested with Landsat data (a 1000 x 1000 pixel image). The results were promising in that tightly-defined classes were defined that were related to identifiable features of the scene, but not rigorously evaluated with ground truth. Also, because the program was computationally intensive, the analysis took many hours on a massively parallel computer, a Thinking Machines Corporation Connection Machine (Model 2). A newer, faster version of AutoClass written in C was tested on a moderate speed platform, a Sun Ultra 30. Software was written to convert imagery in a standard format used in remote sensing into the binary format used for Auto-Class input. Autoclass-C was modified to classify image output in standard formats, enabling map product development and assessment with commercial remote sensing image analysis software. It was further modified to write class parameters in an ASCII file to enable maximum likelihood (ML) classification of a full image with classes defined by AutoClass on a pixel subset. The clustering/ML combination was implemented to allow AutoClass to be applied to mapping larger data sets.