The global illumination in real time of natural scenes such as a forest is one of the most complex problems to solve because of the multiple inter-reflections between light-matter of the objects composing the scene, the major problem that arises is the problem of calculating visibility, indeed the calculation of visibility is done for all the leaves visible from a given point. Given the huge number of leaves present in a tree, this calculation is performed for each leaf of the tree which also reduces the performance.In this thesis we describe a new approach to approximate visibility queries that proceeds in two steps: the first step is to generate a point cloud representing the leaves, we assume that the point cloud is composed of two classes (visible, non-visible) non-linearly separable, the second step is to perform a classification of point cloud by applying the Gaussian radial based kernel function that measures the similarity in terms of distance between each leaf and a kernel leaf.