The main objective of the technique of image data fusion is combining or amalgamating information from multiple sensors such that the new images are more suitable for the purpose of the computer processing tasks such as segmentation

feature extraction

and object recognition. This paper proposes a multisource image fusion algorithm based on hierarchical markov random field model. The algorithm defines the fused image as the hidden labels

and the multisource image can be defined as observation of the hidden labels. The hidden labels can be extended in the quadtree manner. In order to reduce the computational burden

this paper uses a hybrid structure which combines a spatial grid of a reduced size at the coarsest level with sub tree below it

down to the finest level. The hidden labels can be estimated by a noniterative inference on sub tree with ICM algorithm defined on the top of the hybrid structure. In order to circumvent the drawback of classical MAP criterion on the hierarchical graph structures

this algorithm combines nonlinear causal Markov model defined on hierarchical graph structures with bayesian SMAP criterion. This algorithm can applied to multi source remote sensed image fusion

and can contribute to the correctness of image pixels labeling and the reduction of the multi source image volume. The contrast experiments on synthetic and satellite images indicated the advantage of this algorithm relative to the classic algorithm.