In this paper, an architecture for multisensor data fusion and evidence accumulation for landmine detection and discrimination is presented. Evidential and discriminatory information about the buried object such as shape, size, depth, and material, chemical or electromagnetic properties is obtained from different sensor and sensor algorithms. A streamlined assimilation of these varied information from dissimilar and non-homogenous sensor and sensor algorithms is presented. Information theory based pre-processing of the data and subsequent unsupervised clustering using Dignet architecture is used to capture the underlying structure of the information available from different sensors. Sensor information is categorized into type, size, depth, and position data channels. Each sensor may provide one or more of this information. Type data channel provides any relevant discriminatory characteristics of the buried object. A supervised feed-forward neural network is used to learn the causality between the cluster information and the evidence of a given class of the buried object. Size, depth and phenomenology input are used as control gating input for the neural network mapping. The supervisory feedback is provided by the output of the global sensor fusion system and accommodates both autonomous and human assisted learning. Dempster-Shafer evidential reasoning is used to accumulate different evidence from sensor channels and thus to detect and discriminate between different types of buried landmine and clutter. Performance of fusion architecture and Dempster-Shafer reasoning is studied using simulated data. For the simulated data noisy images of regular and irregular shapes of different objects are produced. Fourier descriptor, moment invariant and Matlab shape features are used to define the shape information of the objects. Evidence accumulation is done using shape and size information form each of the algorithms.