FEATURE SELECTION METHODS FOR OBJECT-BASED CLASSIFICATION OF SUB-DECIMETER RESOLUTION DIGITAL AERIAL IMAGERY A. S. Laliberte a , D.M. Browning b , A. Rango b a Jornada Experimental Range, New Mexico State University, Las Cruces, NM 88003, USA – alaliber@nmsu.edu b USDA-Agricultural Research Service, Jornada Experimental Range, Las Cruces, NM 88003, USA – (dbrownin, alrango)@nmsu.edu KEY WORDS: object-based image analysis (OBIA), feature selection, high resolution, aerial photography, vegetation, classification, accuracy ABSTRACT: The availability of numerous spectral, spatial, and contextual features renders the selection of optimal features a time consuming and subjective process in object-based image analysis (OBIA). While several feature selection methods have been used in conjunction with OBIA, a robust comparison of the utility and efficiency of approaches could facilitate broader application. In this study, we tested three feature selection methods, 1) Jeffreys-Matusita distance (JM), 2) classification tree analysis (CTA), and 3) feature space optimization (FSO) for object-based classifications of rangeland vegetation with sub-decimeter digital aerial imagery in the arid southwestern U.S. 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, and classification accuracies. For the five sites tested, JM resulted in the highest overall classification accuracies for three sites, while CTA was highest for two sites. FSO resulted in the lowest accuracies. CTA offered ease of use and ability to rank and reduce features, while JM had the advantage of assessing class separation distances. FSO allowed for determining features relatively quickly, because it operates within the eCognition software used in this analysis. However, the feature ranking in FSO is unclear and accuracies were relatively low. While all methods offered an objective approach for determining suitable features for classifications of sub-decimeter resolution aerial imagery, we concluded that CTA was best suited for this particular dataset. We explore the limitations, assumptions, and appropriate uses for this and other datasets. 1. INTRODUCTION The selection of appropriate spectral bands or image features is a crucial step in any image analysis process. Using a set of op- timal features ensures that the classes in question are discrimi- nated effectively and with sufficiently high accuracy, and that the dimensionality is reduced for efficient use of training sam- ples (Jensen, 2005). In object-based image analysis (OBIA), the determination of optimal features can be a time consuming process due to the availability of numerous spectral, spatial, and contextual features. Feature selection techniques range from graphic methods to statistical approaches involving class sepa- ration distances. Several feature selection methods have been used in conjunction with OBIA. Herold et al. (2003) and Car- leer and Wolfe (2006) used the Bhattacharyya distance, while Nussbaum et al. (2006) and Marpu et al. (2006) employed the Jeffreys-Matusita distance for feature selection. Johansen et al. (2009) evaluated feature space plots, box plots, band histo- grams, and feature space optimization. Classification tree anal- ysis for selection of optimal features was successfully applied by Chubey et al., (2006), Yu et al., (2006), Laliberte et al. (2007), and Addink et al. (2010). The above mentioned studies all used high resolution satellite images (QuickBird, Ikonos, SPOT) or aerial photography (0.3- 1.25m resolution). In recent years, the use of digital mapping cameras has greatly increased, and examples for use of these images include mapping benthic habitats (Green and Lopez, 2007), land use/land cover mapping (Rosso et al. 2008), and border monitoring and change detection (Coulter and Stow, 2008). Digital airborne imagery can be acquired at sub- decimeter resolution and exhibits great potential for mapping rangeland vegetation mapping (Laliberte et al., in press) despite multiple challenges, such as high spatial frequency, the effect of shadows, viewing geometry, illumination, and the necessity for mosaicking multiple images for analysis. Optimal features for classification may be scale dependent, and features used in the analysis of coarser resolution imagery may not be applica- ble to finer resolution data. Determination of appropriate fea- tures for very high resolution imagery, and a robust comparison of the utility and efficiency of various feature selection methods could facilitate broader use of sub-decimeter aerial imagery for vegetation mapping. The objectives of this study were to 1) determine the optimal features for fine-scale vegetation mapping, and 2) evaluate three feature selection methods (i.e., Jeffreys-Matusita distance, classification tree analysis, and feature space optimization), in the context of object-based classification of rangeland vegeta- tion with digital aerial imagery with a 6 cm ground resolved distance. Evaluation criteria for the feature selection methods included efficiency and ease of use, ability to rank and reduce features, and classification accuracies. 2. METHODS Study area and images The study sites were located at the Jornada Experimental Range and the Chihuahuan Desert Rangeland Research Center in southwestern New Mexico, USA (32º34’11”W, 106º49’44”N). Average elevation is about 1200 m, and rainfall amounts and distribution are highly variable, with more than 50% of the mean annual precipitation of 241 mm occurring during July, August, and September. Much of the historic semi-desert grass-