Area estimation from discrete and continuous land cover maps

Many global, regional and even national products are derived from relatively coarse spatial resolution products (250m-1000m). Although useful, such categorical maps may suffer from underestimating classes that are spatially scattered or mostly make up only small patches with complex shapes, which hardly ever fully cover the area of a coarse resolution pixel. An alternative is the estimation of class memberships; that is the proportion of each class in every pixel. This study compares the results from discrete maps and class memberships from two commonly used approaches: Random Forest Classification (RF-C), from which the class probability was estimated from 1000 iterations, and Cubist as a member of regression approaches. The study shows that class memberships better estimate the area than discrete maps. For class memberships, Cubist shows higher accuracies and lower area estimation differences than RF-C, but the contrary pattern is shown for discrete maps.