Submitted to: Meeting Proceedings
Publication Type: Proceedings
Publication Acceptance Date: December 17, 2009
Publication Date: June 29, 2010
Citation: Laliberte, A.S., Browning, D.M., Rango, A. 2010. Feature selection methods for object-based classification of sub-decimeter resolution digital aerial imagery. Proceedings of GEOBIA 2010 Conference, June 29-July 2, 2010, Ghent, Belgium. Vol. XXXVIII-4/C7. Technical Abstract: Due to the availability of numerous spectral, spatial, and contextual features, the determination of optimal features and class separabilities can be a time consuming process in object-based image analysis (OBIA). While several feature selection methods have been developed to assist OBIA, a robust comparison of the utility and efficiency of approaches could facilitate broader application. In this study, we tested feature selection methods and assessments of class separability for object-based classifications of arid land vegetation in the southwestern U.S. using sub-decimeter digital aerial imagery with a 6 cm ground resolved distance. Using Definiens Developer software, we evaluated 1) classification tree analysis (CTA), 2) feature space optimization (FSO), and 3) SEparability and THreshold (SEaTH). We assessed strengths, weaknesses and best uses for each approach using the criteria of ease of use, ability to rank and/or reduce input features, ability to obtain feature thresholds and class separation distances, and underlying statistical assumptions. We conclude that CTA is best suited for reducing a large number of input features, either for nearest neighbor classification or for deriving rules for rule-based classification. FSO is most appropriate for determining a suitable group of features relatively quickly, because it operates within the Definiens software. However, the features are not ranked and feature thresholds cannot be determined. SEaTH interfaces well with outputs from Definiens and is most appropriate for ranking input features based on class separation distances for two-class comparisons, but has the disadvantage of assuming normality, unlike the non-parametric CTA. All methods offered an objective approach for determining suitable features for classifications of hyperspatial aerial imagery. Limitations, assumptions, and appropriate uses for this and other datasets are discussed.