A general multisource contextual classification model of remotely sensed imagery based on MRF

In this paper, we present a general model for contextual multisource and multitemporal classification of satellite imagery based on the Markov random field (MRF). For many remote sensing applications, the robust interpretation and evaluation of remote sensed images require the use of multiple sources of information available about the scene under consideration. Indeed, data provided by individual sensors is incomplete, inconsistent or imprecise. Additional sources may provide complementary information and the fusion of multisource data can create a more consistent interpretation of the scene in which the associated uncertainty is decreased and the reliability of analysis results is increased. Also, temporal data from a single sensor can be considered as separate information sources. The combination of multitemporal data sets over the same scene enhances information on changes that might have occurred in the area observed over time. From all these available data, our objective is to extract more information and achieve greater accuracy in assigning images to thematic classes. The multisource classification model employs a pixel by pixel classification technique that defines the interaction between the different sensors in terms of a sum of sensor specific energy function allowed by MRF. Each sensor is associated with a sensor specific reliability factor. The multitemporal classification model requires the integration into the classification process of the temporal information expressed in terms of transition probabilities from one class at time t-1 to another class at time t according to MRF. The contextual classification model deals with the problem of incorporating into the classification process the contextual information expressed in terms of the spatial interaction that exists between one pixel and pixels in the rest of the scene. This interaction is well modelled by MRF via the Gibbs distribution and Potts model. The general model is tested on three data sets for an urban region of Algiers city. The sets of data are multitemporal LANDSAT TM images acquired during 1985, 1991 and 1996. These sets are both considered as multispectral and multitemporal data. The classification performance is studied in terms of the effect of using remote sensing data from different sensors (respectively of including temporal aspects of data) on the punctual and contextual classification process.