Supervised Self-Organizing Map with Classification Uncertainty

The sensitivity and reliability of the classification output is an important subject for image classification. Classification accuracy in an inference process is always less than a desired accuracy in the actual classification process, thus this marginalized difference is considered as an element of uncertainty in the classification results. Failure to recognize uncertainty may lead to erroneous and misleading interpretations. Therefore, this research aims to quantify the uncertainty of the image classification. The Supervised Self-Organizing Map (SSOM) based on the neural network classification, which is a robust approach and improved image classification accuracy, with the synthetic dataset is used to evaluate the classification uncertainty. Monte Carlo simulation technique is applied to assess the reliability of the classification output by focusing on the uncertainty associated with the input data, training data, and the classifier. The results indicates that increasing the levels of noise have an extensive influence on the classification accuracy. SSOM with different sequences of training data produces the variation of classification accuracy. The minimum number of competitive layer neuron (NET) should correspond to the number of land cover diversities. Initial learning rate (LR) value depends on diversity of study area and the complexity of the input data. SSOM is likely to produce low accuracy and high uncertainty in areas of heterogeneity and large diversity. These results enhance the conceptual understanding of the uncertainty in classification accuracy and the results can also be a guideline to configure appropriate configuration of SSOM to improve classification result.