Multi-Source Image Classification II: An Empirical Comparison of Evidential Reasoning and Neural Network Approaches

Derek R. Peddle,Giles M. Foody,Airong Zhang,Steven E. Franklin,Ellsworth F. LeDrew

DOI: https://doi.org/10.1080/07038992.1994.10874582

IF: 2.6

1994-12-01

Canadian Journal of Remote Sensing

Abstract:SUMMARYConventional and established approaches to image classification, such as the maximum likelihood algorithm, are not always suitable for handling the volume and diversity of modern, multi-source image data sets. In this paper, Dempster-Shafer evidential reasoning (ER) and two Artificial Neural Network (ANN) classifiers are explored as alternative techniques to maximum likelihood (ML) for multi-source image classification. Each technique is tested in separate classifications of alpine land cover and permafrost active layer depth for a mountainous area in the southwest Yukon Territory using a data set of SPOT imagery, image texture, and geomorphometric variables from a Digital Elevation Model. In comparisons of average overall accuracy, the ER classifier was 10% higher than ML and the ANN classifiers were 5% higher than ER. The highest overall kappa co-efficients of agreement obtained for land cover were 0.79, 0.90, and 0.96 for ML, ER, and ANN, respectively. Substantial differences in accuracy between the two ANN classifiers were noted in several instances; these were attributed to difficulties in specifying optimal parameters for neural network training from subjective user input. In terms of computing time, the evidential reasoning algorithm was significantly faster than the neural network classifiers by several orders of magnitude. From these experiments, we conclude: both the ER and ANN classifiers are superior to maximum likelihood for handling multi-source data; evidential reasoning is faster and less subjective compared to neural networks, but ANN provides higher accuracy and is therefore the recommended procedure at present — this is what we expected given the maturity of neural network research compared to evidential reasoning, which is a relatively new technique in its early developmental phase in terms of remote sensing applications; further research is warranted for the continued development of evidential classifiers and also to determine optimal user input parameters for neural network classifiers to improve its level of objectivity; and, evidential reasoning and neural network approaches appear well suited for multi-source image analysis and classification of complex environmental phenomena.

remote sensing