Many works dealing with the problem of urban detection in large scale have been
published, but very little attention has been paid  to the investigation of the
features relative importance. Feature selection is known to be an NP-hard
problem, with many heuristics suggested to approximate the solution.
In this paper, a quick survey of the features used for large scale urban detection using Landsat
data is presented, then the question of finding the best subset of features is
investigated. Using Landsat scenes of five urban areas, all common features were
extracted to represent the full feature set. Employing mutual information based ranking methods,
Fisher score, SVM and Random Forest feature ranking, an importance score was
assigned to each feature by each method. To aggregate the individual rankings of features, a
two stage voting scheme was implemented to choose a subset of size $N$
 as the most relevant features.
To evaluate the chosen subset, a comparison to a baseline subset was
performed. The classification power of the two subsets was tested using four
classifiers in five urban regions. The results suggest better performance of the chosen subset
compared to the baseline.