Spatial outlier detection has been a hot issue in the field of spatial data mining and knowledge discovery. Spatial outliers may be utilized to discover and predict the potential change laws or development tendency of geographical phenomenon in the real world. Among the existing spatial outlier detection methods

there are mainly two aspects of issues. On the one hand

these methods primarily consider that all the entities for outlier detection are correlated. Actually

spatial correlation decreases with the increase of distance. Entities will become independent with each other at a distance of rang. Thus

current methods can only discover the obviously outliers in the whole

some local outliers may not be detected. On the other hand

the spatial outlier measures are not enough robust

which are seriously influenced by the construction process of spatial neighborhoods of spatial entities and the possible outliers in spatial neighborhoods. To overcome these two limitations

spatial clustering as a means is firstly employed to extract the local autocorrelation patterns

called clusters. Then

a robust spatial outlier measure is proposed to determine spatial outliers in each cluster. This method is able to detect spatial outliers more accurately. Finally

a practical ex-ample is utilized to demonstrate the validity of the spatial outlier detection method proposed in this paper. The comparative experiment is also provided to further demonstrate the method in this paper to be superior to classic SOM method.