Random forests is a widely used classification algorithm. It consists of a set of decision trees each of which is a classifier built on the basis of a random subset of the training data-set. In an environment where the memory work-space is low in comparison to the data-set size, when training a decision tree, a large proportion of the execution time is related to I/O operations. These are caused by data blocks swap-in from the storage device to the memory work-space. In traditional methods, data blocks contain elements that will be accessed and processed at different moments, thus, the spatial locality is low. In this paper, we seek to reduce the number of data-blocks swap-in operations by enhancing spatial locality. The idea is to reorder dataset blocks such as each block contains elements that are likely to be accessed together. Our experiments show that this method reduces random forest build time by 55% to 89%.