Feature Selection Methods for Object-Based Classification of Tropical Forest Using Spot 5 Imagery

Abstract

The faster development of spatial technology, the more options in material for image analysis and this characteristic also leads to a huge diversity of image features. Aiming at investigating the optimal feature space for classification with emphasis on efficiency in less time-consuming, image objects separability, certain numbers of approaches ranging from graphical to statistical methods involving class separation distances have been used for classification related to Object-based Image Analysis (OBIA). Hence, a robust comparison of the utility and efficiency among methods is necessary. This paper presents the comparison between two approaches of optimal feature selection in the context of forest classification with Spot 5 imagery in the Nongha commune, the North eastern Vietnam. The first one is Jeffreys-Matusita distance, which uses the tool named Separability and Threshold (SEaTH) running in an IDL virtual machine environment and the other is Euclidean distance using the feature space optimization tool in eCognition software (FSO). This study mainly focuses on assessment of processing time, simplicity in using, feature space reduction and classification accuracy. After testing, SEaTH method shows advantages in term of classification accuracy, while FSO is a quite simple method which is user-friendly and does not require much in complicated manipulations. Finally, each approach results in different advantages and disadvantages. In the context of this study, which emphasize on forest classification using Spot 5 image, it is possible to conclude that SEaTH is the best choice.

Keywords: Object-based Image Analysis (OBIA), Classification, Forest types, Separability and Threshold (SEaTH), Feature Space Optimization (FSO)