Remote sensing is the set of techniques that allow, by means of sensor technologies, to analyze objects at long distances without making physical contact with them. Currently, its contribution for natural sciences is enormous, since it is possible to acquire images of target objects in more regions of the electromagnetic spectrum than the visible region only. Working with images composed of various spectral bands demands dealing with huge amounts of data associated with single entities, which affects negatively the performance in prediction tasks, and makes necessary the use of dimensionality reduction techniques. This work introduces a feature extraction approach, based on spectral indices learned by Genetic Programming (GP), to project data from pixel values into new feature spaces aiming to improve classification accuracy. Spectral indices are functions that map the reflectance of remotely sensed objects in specific wavelength intervals, into real scalars that can be interpreted as the abundance of features of interest. Through GP, it is possible to learn indices that maximize the separability of samples from two different classes. Once the indices specialized for all the pairs of classes are obtained, they are used in two different approaches to fuse them into a pixel classification system. Results for the binary and multi-class scenarios show that the proposed method is competitive with respect to traditional dimensionality reduction techniques. Additional experiments in tropical biomes seasonal analysis show clearly how superior GP-based spectral indices are for discrimination purposes, when compared to indices developed by experts, regardless the specificity of the problem.