A Classification Method Based on Spatial Information

赵红蕊,阎广建,邓小炼,王锦地,杨华,李小文
DOI: https://doi.org/10.3321/j.issn:1007-4619.2003.05.004
2003-01-01
Journal of Remote Sensing
Abstract:Remote sensing image classification is one of the basic components in remote sensing image processing, which is the key stage in remote sensing applications. There are two ways to get a higher precision: one is to make full use of spectral information, and another is to add other new information The later is called multi-information processing. Spatial information is one of the most important information in these kinds of multi-information. Since lack of information is always the key problem in remote sensing image processing, we propose a spatial information based method to cope with this problem. What we used is not the spatial data but the spatial characteristics. The main advantage of our method is that it can work without any other data sources. We take advantage of spatial autocorrelation law, which is the first law of geography. The law is widely used in geo-analyses. It denotes that the closer in spatial, the similar the objects or phenomenon is. Two spatial bands are constructed based on spatial autocorrelation law and a spatial classification result is produced. These two additional bands containing spatial location information are used in classification. One real class may be divided into many classes that are not similar in space. But every class has similar spectral and spatial characteristics. Based on the classification results, further processing should be done. First, we replace every pixel value of every spectral band with the average spectral value of its class. Then a traditional classification method, such as ISODATA method, can be used. With these steps, we get the last classification results. The MAS and ASTER images are used to prove our method. The classification results show that some miss classified ground covers in the general spectrum based classification method are thrown off by our method. In other words, our method is resistant to the problem of having apparent same spectrum but different ground cover. Another distinct advantage of our method is about the removal of salt and pepper noise from images. It may be also helpful in other research area, such as landscape research. Our method also proves the importance of spatial information in remote sensing image classification. It gives an easy way to add spatial constraint to classification process. The new classification method reveals its simpleness and practicality.
What problem does this paper attempt to address?