Printer friendly version
Major step towards the application of self-organizing neural networks for remote sensing
23 March 2010
Universidad Politécnica de Madrid
Research developed at the Universidad Politécnica de Madrid's School of Computing has brought us a big step closer to the application of self-organizing neural networks in the remote sensing field.
To be more precise, this research has developed new self-organizing neural network training and visualization algorithms for application in remote sensing, generating simplified models of large volumes of multi-spectral information.
Neural networks are mathematical models inspired by the operation of biological neural networks. They are now applied across a wide range of disciplines to solve a broad spectrum of problems. One of the most widely used neural network models is what is known as the self-organizing map.
Remote sensing is a discipline that is concerned with the acquisition of information about the Earth's surface without entering into physical contact with the object under observation. The development of tools for analysing and processing multi-spectral images captured by sensors on-board satellites has paved the way for automating tasks that would otherwise be impracticable.
Large data volume
The key problem related to remote sensing is the large volume of multi-dimensional data that has to be managed. The self-organizing neural network, and specifically Kohonen's model, has proved to be a versatile and useful tool for exploratory data analysis.
But Kohonen's model has some, primarily architecture-related, constraints. This has led to the emergence of new types of self-organizing maps, like the Growing Cell Structures (GCS) model, that tackle these issues.
Using the GCS model, the relationships of the information input patterns can be visualized without the topological constraints of Kohonen's model. On the downside, though, some training parameters are hard to configure, as, even if constant values are assigned, it is not clear what the permitted value range for these patterns is.
By proposing a new GCS model training algorithm that improves this network-input space topology fit, the research developed at the School of Computing comes closer to solving this problem.
The modification of the GCS algorithm makes this neural network model easier to use to generate simplified models of the large volumes of multi-spectral information typically associated with the remote sensing field.
With the aim of exploiting this paradigm within the remote sensing field, the research project has developed several GCS-based multidimensional information visualization methods, as well as a number of network labelling techniques for semi-supervised and unsupervised classification or multispectral information-based variable estimation processes. Additionally, as part of the research, several GCS-adapted measures have been developed to evaluate the quality of the trained network.
The developed methodology has been applied across a range of hot topics in the remote sensing field, like classification of land covers in semi-supervised and supervised processes, evaluation of the quality of training areas selection, estimation of the physical variables of aqueous covers or the analysis of spectral index validity for images with special features.
Application to other areas
The features of the developed tools make the proposed methodology a very useful tool in other research fields that, like remote sensing, have to manage multidimensional information.
For this reason, this research has included experiments related to DNA strand management, as well as the processing of medical data related to kinematic variables of walking gait in children. This served to validate the developed methodology.