Object-oriented feature selection of high spatial resolution images using an improved Relief algorithm

In object-oriented classification approaches, small sample size and high dimensionality of features are the two main characteristics. In order to effectively use the rich features of the image objects, in this study, the Relief algorithms were improved in terms of aspects of randomly drawing samples, the influence of sample quantity variance, and iteration times to evaluate the features. Through experiments in WorldView-2 images, the two classification results were contrasted between 36 dimension feature sets selected from Initials and initial 111 dimension feature sets. The results showed that the dimension disaster was avoided efficiently, while the precision and speed of classification were improved as well. Respectively, the overall accuracy and kappa coefficient using the optimized feature sets with the improved Relief algorithm were increased by 6% and 11% compared with using all feature sets. Therefore, the feature space of object-oriented classification is effectively optimized with the proposed methods.