Using artificial neural networks for synthetic surface fitting and the classification of remotely sensed data

Abstract The potential of neural networks in multidimensional interpolation was explored using a multilayer perception for fitting surfaces to a synthetic topographic dataset. Density-sliced and shaded relief images generated through this neural network-based surface-fitting scheme were compared with those generated by conventional approaches, eg, Akima's quintic polynomial fit and inverse square method [Akima, 1978]. Compared with the conventional approaches, the neural network approach was found to better represent the nonlinearity in the synthetic dataset. This paper also presents a method for classifying remotely sensed data, using an artificial neural network (ANN) approach. The ANN used was a multilayer perception trained through the generalized delta learning rule. The software package was completely generalized in nature and could deal with any number of input units (spectral bands), output units (feature classes) and hidden layers. Different numbers of hidden neurons could also be considered in various hidden layers. The software package was also used for classifying IRS-1A LISS-1 images.