Which are the most common features in order to perform a supervised classification in remote sensing?

Furthermore anyone worked with SAR features?

Which are the most commonly used?

Can anybody suggest me a book for feature extraction?

  • could you please be more specific concerning the type of classification output that you expect, the scale of your work and the data that you have access ? as a first answer, I would say that the most "COMMON" features are near-infrared and visible wavelength. But advanced features can be helpfull depending on your specific tasks.
    – radouxju
    Commented Nov 12, 2014 at 9:16
  • rgb false color image IS a multispectral data.
    – radouxju
    Commented Nov 12, 2014 at 10:39

2 Answers 2


You should provide more details about the task. In general, extracting features from the data image heavily depends on what you are trying to detect/classify, and how are you trying to do it.

Here's an example. If you are interested in classifying roads from an urban scene, you may be interested in evaluating large linear filter responses over the whole scene, to extract some signal about elongated and contiguous elements. If you aim at segmenting homogeneous regions, e.g. cultivated crops, you may want to compute features reducing the local class variance, e.g. local means, to reduce the within class variance and thus improve the between-class separation. Another classical example is to extract texture features (e.g. Haralick's GLCM features, local standard deviation, etc) to improve discrimination between spectrally similar surfaces but with different textures / pixel values distribution.

In general, it is common to stack these extracted features to the original image, so that the spectral information (either RGB, NIR, hyperspectral, whathever it is) is complemented with task specific information. In general, this is not limited to supervised classification, but large improvements are observed by using such models. Be just aware that by stacking many features to the original image the dimensionality may drastically increase. Think for instance stacking multiscale version of the examples above (say 3x3, 5x5, ..., 25x25) to an hyperspectral image. The dimension may become too large for some classifiers in particular in small sample size scenarios, so be careful.

Regarding SAR images, I'm no expert. However, I'm pretty sure that the big deal is to compute filterings of the original data, so that noise is reduced. I guess that you can simply transfer the knowledge about features / descriptors from optical images to SAR images after a log transform to make noise additive. (I'm sure that the truth is far from this simplification).

To have a feeling with more recent state of the art and literature reviews, I suggest to browse remote sensing journals (IEEE TGRS, IEEE GRSL, IEEE JSTARS, ISPRS JPRS, RSE, etc) and search for "feature extraction", "spatial filterings", or for specific families of filters, e.g. "GLCM", "texture", "mathematical morphology". Be careful since the term "feature extraction" may also refer to its "machine learning" variant, which related to rotation-based transformation (e.g. principal component analysis) and not directly related to the extraction of "meaningful" information from images. Though in both cases the aim is to facilitate the analysis of the data...


If you want to know basics about the remote sensing, here is the link which might be helful for you, http://grindgis.com/what-is-remote-sensing/know-basics-of-remote-sensing

  • This is essentially a link-only answer, which is generally discouraged here on GIS.SE. You should at least provide a summary of the information found at the link. Links can go dead over time, rendering the answer useless. Also, this question is specific to feature extraction - does your link, which seems to indicate a generic primer on remote sensing, specifically cover that topic?
    – Chris W
    Commented Jun 4, 2015 at 1:23

Not the answer you're looking for? Browse other questions tagged or ask your own question.