Research work at the School of Computing of the Universidad Politecnica de Madrid has reached a milestone that is close to the development of self-organizing neural networks for use in the remote sensing domain.
This research work has resulted in the development of visualization and training algorithms for self-organizing neural networks. The algorithms can be used in remote sensing applications and also evolve simple models that would help to collect large amount of multi-spectral information.
The functioning of the biological neural networks has triggered the emergence of neural networks that are basically mathematical models. Such models are used for a multitude of disciplines for solving a wide range of problems. The self-organizing map is among the mostly utilized neural network models.
The remote sensing domain deals with information acquisition related to the surface of the Earth without contacting the observed object physically. Tools developed to process and analyze multi-spectral images collected by on-board satellite sensors has resulted in the automation of tasks that otherwise would not have been possible.
The management of a huge amount of multi-dimensional data poses a major hurdle for remote-sensing applications. The Kohonen model of self-organizing neural networks has been proved as a useful and versatile tool for the analysis of exploratory data.
However, this model has basic architectural-oriented drawbacks, which has resulted in the emergence of newer forms of self-organizing maps, such as the growing cell structures (GCS) model, for countering these drawbacks. The GCS model can be used to visualize the relationships among information input patterns without constraints imposed by Kohonen’s model. On the minus flip side some of the parameters related to training are difficult to configure. Even after assigning constant values the allowed value range for these information patterns is still unclear.
To exploit this concept in the remote sensing domain several GCS-enabled multidimensional information visualization processes have been developed during the research project, besides numerous network labeling methods for unsupervised and semi-supervised classification. As part of the research project, in addition, several GCS-enabled measures have been created for the evaluation of the trained network quality.
The developed process has been used for a range of critical topics in the remote-sensing domain, for instance, categorizing land covers for supervised and unsupervised methods, estimating physical variables of aqueous covers, analyzing the validity of spectral index for special features and images.
Characteristics of tools developed render the process a utility tool that could be used in various research fields that require the management of multidimensional information, such as remote sensing.
The research project has DNA strand management-related experiments, and medical data processing associated with kinematic-based variables in walking children. This helped researchers to authenticate the methodology developed.