Neurocomputational Spatial Uncertainty Measures

Both the Self-Organizing Map (SOM) and fuzzy ARTMAP neural network are trained based upon the competitive mechanism and use the “winner-take-all” rule. Previous studies developed soft classification algorithms for the SOM. This paper introduces the idea and proposes non-parametric measures for the fuzzy ARTMAP computational neural networks to handle spatial uncertainty in remotely sensed imagery classification. These soft algorithms are neuron-triggering/committing-frequency based and are grouped into two types, i.e., Commitment and Typicality, expressing in the first case the degree of commitment a classifier has for each class for a specific pixel and in the second case, how typical that pixel’s reflectances are of the ones upon which the classifier was trained for each class. Two measures are designed for each of the two neural network models, i.e., SOM Commitment (SOM-C) vs. SOM Typicality (SOM-T) and ART Commitment (ART-C) vs. ART Typicality (ART-T). To evaluate these proposed algorithms, soft classifications of a SPOT HRV image around Westborough, Massachusetts were undertaken. Conventional soft classifiers such as Bayesian posterior probability classifier and Mahalanobis typicality classifier were used as a comparison. Correlation Analysis and Principal Components Analysis (PCA) were employed to explore the relationship between these different measures. Results indicate that great similarities exist among the ART-C, SOM-C and the Bayesian posterior probability classifier, and significant similarities exist among the ART-T, SOM-T and the Mahalanobis typicality classifier. However, the proposed models outperformed conventional ones.