A Novel Multi-radius Neighborhood Rough Set Weighted Feature Extraction Method for Remote Sensing Image Classification

The neighborhood rough set model can be effective for keeping or even improving classification accuracy. This model however, still has some disadvantages as it has low stability in classification precision, requires repeated neighborhood radius adjustments, and cannot realize automatic feature extraction. In order to solve these problems, this paper presents a multi-radius neighborhood rough set weighted feature extraction method for high resolution remote sensing image classification. The neighborhood rough set model was used to extract texture and spectrum features of image by setting gradually increasing radius with equally spaced steps, as a result effective subset features under different radius were obtained. The presence probablity of each feature under all the different radius was calculated, each feature was endowed with weight by its presence probability, so the final weighted features of image were acquired. The newly obtained features were applied to image classification using a support vector machine. Experiments on QuickBird images demonstrate that the proposed method can provide better classification results. Compared with other state-off-art neighborhood rough set model with effective radius, the overall accuracy exceeded about 3.88% while the Kappa coefficient exceeded about 5.16%. A classification experiment on a GeoEye-1 image also showed the effectiveness of the proposed method. All the classification experiment results show that the proposed method can improve classification precision and automation of high resolution remote sensing images.