Low Classification Accuracy for Hyperspectral Imagery? Try Attribute Profiles!



Recent advances in hyperspectral remote sensor technology allow the simultaneous acquisition of hundreds of spectral wavelengths for each image pixel. This detailed spectral information increases the possibility of more accurately discriminating materials of interest. Further, the fine spatial resolution of the sensors enables the analysis of small spatial structures in the image.

The need of spatial structures in the classification

The challenge for the classification usually is to distinguish the spectrally similar classes such as soil-roads-roofs, trees-grass, and water-shadow, which means we should propose different methods to efficiently extract the different features.

However, the most available hyperspectral data processing techniques usually analyze the data using spectral information without particular spatial arrangement. In certain applications, Researchers have identified it is important to use both the spectral and spatial information.

“Small spectral classes need to be identified with sufficient spatial resolution in urban area mapping (Gamba et al. 2014)”, which means processing algorithms should take spatial information into account.

How to integrate spatial signature with spectral information?

To address the need for knowledge-based developments, and take into account both spectral and spatial properties of the data, researchers have used attribute profiles to extract information about the size, shape and the orientation of structures in single-band remote sensing images. “Improved attribute profiles can deal with the full spectral information available in the data (Marpu et al. 2013).”

In addition to mathematical morphological-based approaches, Markov Random fields (MRFs) can also be used to model the spatial neighbourhood of a pixel as a spatially distributed random process, and attempt regularization via the minimization of an energy function.

Why Attribute Profiles?

They either completely remove or entirely preserve a structure in the image, and do not distort shape of structures nor introduce new edges. So they are suitable for the analysis of very high resolution images.

They are flexible tools: attributes can be defined in any way. For instance, they can be purely geometrical (e.g., area, moment of inerita) of related to the gray-scale distributions of the pixels in the regions (e.g., std., entropy, uniformity, contrast)

The union of attribute profiles and image representation leads to an efficient and fast procedure for the computation.

The use of a feature extraction technique leads to a further increase in terms of accuracies when compared to the use of the data with full dimensionality.

Application to specific tasks such as object detection (e.g., building detection, road networks extraction) and multi-temporal image analysis (e.g., including the modelling of spatial information provided by attribute profiles in the change detection analysis).

The procedure of classification using Attribute Profiles

In the figure, AP stands for Attribute profile


Mura, M.D., Benediktsson, J.A. and Bruzzone, L., 2010, July. Classification of hyperspectral images with extended attribute profiles and feature extraction techniques. In Geoscience and Remote Sensing Symposium (IGARSS), 2010 IEEE International (pp. 76-79). IEEE.


There is a limit for attribute profiles which is when using operators based on structuring elements, each threshold used by the criterion needs to entirely process the image. Maybe reducing the dimensionality of the data to few significant bands and applying the operators on each of them is a solution.


Recent Papers related to this topic:

Camps-Valls, G., Tuia, D., Bruzzone, L. and Atli Benediktsson, J., 2014. Advances in hyperspectral image classification: Earth monitoring with statistical learning methods. Signal Processing Magazine, IEEE31(1), pp.45-54.

Fauvel, M., Tarabalka, Y., Benediktsson, J.A., Chanussot, J. and Tilton, J.C., 2013. Advances in spectral-spatial classification of hyperspectral images. Proceedings of the IEEE101(3), pp.652-675.

Ghamisi, P., Dalla Mura, M. and Benediktsson, J.A., 2015. A survey on spectral–spatial classification techniques based on attribute profiles.Geoscience and Remote Sensing, IEEE Transactions on53(5), pp.2335-2353.

Li, J., Huang, X., Gamba, P., Bioucas-Dias, J.M., Zhang, L., Atli Benediktsson, J. and Plaza, A., 2015. Multiple feature learning for hyperspectral image classification. Geoscience and Remote Sensing, IEEE Transactions on53(3), pp.1592-1606.


Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out /  Change )

Google+ photo

You are commenting using your Google+ account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )


Connecting to %s