5 February 2004 Optimal design of neural networks for land-cover classification from multispectral imagery
Author Affiliations +
Abstract
It has long been shown the effectiveness of artificial neural networks to solve highly non-linear problems such as land-cover classification based on multispectral imagery. However, due to the large amount of data that is processed within this kind of applications, it is desirable to design networks with the lowest number of neurons that are capable to separate all of the given classes. At present, there are several methods intended to determine this optimal network. Most of them involve adjoining or pruning hidden neurons followed by further training in iterative fashion, which is generally a very slow process. As an alternative, the approach described in this paper is based on the computation of centroids of relevant clusters for each class samples through the well known clustering method ISODATA. A proper tessellation of the ISODATA centroids allows first the determination of the minimum number of neurons in the first hidden layer that are required to effectively separate all of the classes; and secondly, to compute weight and bias parameters for such neurons. Then, the minimum network required to perform the logic function that combines the halfspaces generated by the first layer into class-discriminant surfaces is determined via a logic function reduction method. This approach is much faster than that of current methods because it allows to determine the optimum network size and compute weight and bias parameters without further iterative adjustments. The procedure was tested with landsat 7 Enhanced Thematic Mapper Plus (ETM+) data. Results indicated that (1) the network exhibits good generalization behavior and (2) classification accuracies do not depend on the class boundary complexity but only on the class overlapping extent.
© (2004) COPYRIGHT Society of Photo-Optical Instrumentation Engineers (SPIE). Downloading of the abstract is permitted for personal use only.
Jose L. Silvan-Cardenas, "Optimal design of neural networks for land-cover classification from multispectral imagery", Proc. SPIE 5238, Image and Signal Processing for Remote Sensing IX, (5 February 2004); doi: 10.1117/12.513981; https://doi.org/10.1117/12.513981
PROCEEDINGS
12 PAGES


SHARE
Back to Top