Applying machine learning to Big Data can be a cumbersome task which requires a lot of computational power and memory. In this paper we present a feature selection technique for land-cover classiffication in earth observation scenario. The technique extends the state-of-the-art feature extractors by pruning the dimensionality of the required feature space and can achieve almost optimal results with 10-fold reduction of the number of features. The approach utilizes a genetic algorithm for generation of optimal feature vector candidates and multi-objective optimization techniques for candidate selection.