The sensitivity of a neural network for classifying remotely sensed imagery

Abstract

A series of experiments are conducted on a feed-forward backpropagation neural network which is used to classify land cover from Landsat TM data. By investigating the effects of changing the numbers of network nodes in the input and hidden layers, potentially surplus nodes can be identified and removed to create a more compact network, without loss of classification accuracy. By exploring how momentum can be used with different rates of network learning, an optimal pairing is found which leads to a more rapid convergence and better classification of urban land cover than obtained in previous studies where momentum rarely was used. These optimal network parameters are used to classify an extract of a Landsat TM image of a dockland area with accuracy equal to that obtained using the maximum likelihood method. Given that in this case, the nature of the image data is ideal for a parametric method, this result is not unexpected. The competence of the neural technique is however demonstrated and criteria are given to help determine in advance when neural techniques may be preferable to parametric classifiers. Taken together, the findings show that careful balancing and adjustment of network parameters may be required to obtain a satisfactory result. The method can guide new users in configuring a popular neural network to suit their image data. Given the specific nature of our results, further research on neural networks in remote sensing could benefit from more systematic reporting of network parameters, training times and accuracies obtained.