The spatial and spectral information in remote sensing data should be exploited to improve classification precision.Unfortunately

the spatial information is neglected in most traditional remote sensing classification methods.Considering two stages probability method of maximum likelihood classification（MLC）

this article proposed a new method of exploiting spatial information to improve classification rules by adjusting the prior probability according to the local spatial information.The local or global spatial features of typical ground targets were analyzed

and there were four principles proposed to decide the prior probability

including spatial geometrical features

local spatial autocorrelation law

contextual knowledge

and landscape parameters.An algorithm was designed to exploit the spatial features and prior knowledge to adjust the prior probability of MLC.The experimental classification was carried out

the classification error matrices and its precision results showed that the floating prior probability MLC method proposed in this article could integrate ground targets’ spatial knowledge with spectral information

and overcome the flaws of traditional pixel-based classification methods

such as minimum distance

MLC

and that the classification precision was improved greatly.