A Novel Wrapper Approach for Feature Selection in Object-Based Image Classification Using Polygon-Based Cross-Validation

Abstract

Feature selection is becoming a major component of object-based classification as numerous features of segmented object become available. Although common feature selection methods in object-based classification are acknowledged, wrapper-based methods remain an issue due to the diversity of accuracy assessment methods. This letter presents a new wrapper approach using polygon-based cross validation (CV) to overcome possible bias of object-based accuracy assessment for object-based classification. The new method is a two-step wrapper-based feature selection that involves the integration of: 1) feature importance rank using gain ratio and 2) feature subset evaluation using a polygon-based tenfold CV within a support vector machine (SVM) classifier. Several high-resolution images, including both unmanned aerial vehicle images and ISPRS (International Society for Photogrammetry and Remote Sensing) benchmark test data, were used to test the proposed method. Results show that, with the proposed polygon-based CV SVM wrapper, the mean overall accuracy is significantly higher than with an object-based CV SVM wrapper. Furthermore, the proposed method shows potential for comprehensively considering all types of features instead of only spectral features.