Susana Arias, Héctor Gómez, Francisco Prieto, María Botón, Raúl Ramos
Medio de Publicación
Congreso: EELA2, Choriní (Venezuela)
Año: 25-27 Noviembre 2009
Tipo de publicación: Oral
The present contribution describes the main goals and methodology for the deployment of a self-organized neural network algorithm (SOM) for satellite image classification on a GRID infrastructure. Satellite images are first handled through a normalization process in order to obtain the equivalent data matrices. Then, two phases are implemented to build the image catalogue; training and classification. Each process is first introduced in its local form and then a GRID based solution is proposed.
For the training phase, a collection of selforganized maps with the same input data matrix as codebook vectors are sent to the GRID. Each SOM training process is then handled as an independent job. As a result, a trained SOM and quantization error are obtained. From this set the SOM corresponding to the lowest quantization error is selected. In the second stage, the selected SOM, labelled according to specified classes, is used for the image classification. For this, a GRID scheme is also implemented. Here a set of target images using the previously selected SOM is sent to the GRID. Then the corresponding output is retrieved.
In order to validate the proposed scheme, two use cases were analysed. The target data were images obtained by the Landsat satellite at the Loja and Zamora-Chinchipe regions in Ecuador.