A remote sensing data classification method using self-organizing map

A supervised classification method using a self-organizing map (SOM) is proposed to classify remote sensing data. The SOM structure is composed of two layers. One is an input layer with nodes corresponding to spectral bands. The other is an output layer with square array of nodes. First, a feature map on the output layer is generated by inputting training data into SOM. Each node in the feature map cannot be corresponding to a category because the number of nodes is generally greater than those of training data. Thus, a cluster map is generated by comparing differentials among weight vectors in nodes. Secondly, the training data is re-inputted into the cluster map to find the relationship between clusters and categories, that is, the cluster including a fired node is labeled as the category to which the training data belongs. In consequence of mapping, the category map is obtained from the feature map. The proposed classification method extracts liquefied area in Kobe (Japan) damaged by the 1995 Hyogoken Nanbu earthquake using the SPOT HRV data and the category map. As an experimental result, it is shown that classification accuracies of the proposed method are higher than those of the maximum likelihood and the backpropagation methods