Classification of remotely sensed imagery using adjacent features based approach

The land cover types in urban fringe areas are relatively complex so as to improve the classification accuracy difficultly. This article analyzes the distribution characteristic in feature space of the pixels with a local window from satellite image on a part of SPOT from an urban fringe area in Beijing. There are two methods with different input parameters of using artificial neural networks to describe this distribution characteristic: the input parameters are made up of the spectral information of the pixels in a 3 times 3 window; the input parameters are made up of the spectral information of the center pixel and the statistical distance of the pixels in a 3 times 3 window. After comparison classification results based on the method using adjacent feature, the first method is better than the second method on overall accuracy and kappa coefficient. However, the performance of the first method is lower than the second method in capability of habitation and minimal land objects detection.