Classification of Remote Sensing Data Using Combination of Supervised and Unsupervised Neural Network

In this paper an algorithm for classification of land cover from remote sensing data based on the combination of supervised and unsupervised neural networks is presented. The proposed algorithm composed of Self- Organizing Map (SOM) and Multi-Layer Perceptron (MLP) algorithms. Since the SOM is an unsupervised algorithm, it can’t determine the accurate label of image pixels by itself; therefore, in this paper, the MPL was used to determine the final label. The proposed algorithm includes two steps. First, the image segmented is performed using the SOM algorithm; second, the labels of all SOM’s neurons are determined using MLP and training data to produce the land cover thematic map. In this paper the different numbers of neurons were considered in SOM structure for reducing the number of mathematics operations. In this case, the PCA algorithm was applied for initialization SOM and the result of this part shows that the number of operations was reduced significantly. The algorithm performed on Landsat (ETM+) and Ikonos images to demonstrate its capability. Different combination algorithms were preformed in this work. All these algorithms were used to produce land cover map and then the obtained result were compared with SOM-MLP results. The obtained results showed the ability of SOM-MLP algorithm for land cover classification. In this regard, the result of maximum likelihood classifier, minimum distance algorithm and MLP were compared to proposed algorithms result. Finally, it is concluded that the SOM-MLP improves the accuracy of classification, and it is suitable to remote sensing data when there is not enough ground truth data for training.