Improving Classification Accuracy of High Spatial Resolution Images by Using Texture Quantization and Genetic Feature Selection

Texture quantization is a useful method for extracting spatial relevance between pixels, which is used in the human brain for image interpretation. Aside from spectral bands, textural features of high spatial resolution image can be used to improve classification accuracy. Finding proper textural features among available features is important for special case studies.

In this paper, two methods based on genetic algorithm (GA) are introduced to choose efficient features. The first is binary GA, which improves classification accuracies through selecting the best textural features. The second one is GA with a variable number of selected features in a refined and full feature space. Results show that the best combination does not necessarily consist of features with improved individual accuracy.

The proposed methods have better accuracy, less number of features, and less computational time when comparing with the simple GA. They could be used based on the number of spectral bands, number of generated features, and train and check pixel number. Second method needs more prerequisite time and could be used for images with fewer bands, train and check pixels, and generated features, because increasing these items increase computational time very much. Second method could be used in large images with more train and check pixels but led to more selected features.

Results obtained on three datasets indicate 7.7 to 50.48 percent improvement in mean accuracy.