Classification of remote sensing data by multistage self-organizing maps with rejection schemes

A new classification method for remote sensing data is proposed. The proposed classifier consists of several stage neural networks (SNN) and rejection schemes. Rejection schemes are used to decide whether the input vector is hard to classify. By adopting rejection schemes, it is possible to detect the hard input vectors and reduce the possibility of misclassification, for example, due to input vectors which are linearly non-separable or close to boundaries between classes. Such input vectors are rejected by rejection schemes in each SNN and fed into the next SNN. Simultaneously, the input vectors accepted by rejection schemes are classified in each SNN. The self-organizing map (SOM) is used for learning of weight vectors. Experiments are done using the proposed method with two remote sensing data sets, and results are compared to those of other methods.