A comparison study of selected training algorithms for neural networks with regard to classification of remotely sensed multispectral images
Research output: Contribution to journal › Article › peer-review
Neural networks are increasingly used to analyse and classify multispectral remote sensing images. The most popular neural algorithm in current use is the multi-layer perceptron using some form of back propagation training algorithm. However the multi-layer perceptron is a difficult algorithm to use and very importantly also suffers the drawback of long training times. More recent neural algorithms may prove more suitable for remote sensing users, yet there has been little examination of alternative neural algorithms in the literature and this lack of information makes it difficult for prospective users of neural technology to make an informed choice on the algorithm to use for any application. The following work describes a classification exercise carried out on a multispectral Landsat TM image using three radically different neural networks using the multi-layer perceptron, binary diamond and radial basis function algorithms. The salient features of each algorithm: ease of use, speed and accuracy were examined.
|Number of pages||6|
|Journal||Advances in Space Research|
|Publication status||Published - 1998|