Scale-sets image classification with hierarchical sample enriching and automatic scale selection

Abstract

Object-based image analysis (OBIA) has been widely used for classifying high-spatial-resolution images, and the selection of scale parameter(s) is inevitable in previous OBIA tasks. However, selecting appropriate scale(s) for an application inherently depends on the objective of the application, which cannot be robustly solved at the segmentation stage. In this study, a novel framework without any scale parameter was proposed for object-based image classification. It consists of four major steps: (1) Scale-sets image representation: Multiscale segments from a region-merging algorithm are organized using a scale-sets structure, where each segment is represented as a node of the hierarchy. Note that the implementation of this step does not need any scale parameter. (2) Multiscale sample selection and enriching: Multiscale segmentation results are retrieved from the scale-sets structure and visualized, and then training samples are selected from multiple scales. The training samples are further enriched according to hierarchical relations. (3) Feature extraction and multiscale classification: Segments are described using spectral, textural and geometric features, and then classified using a Random Forest classifier. Through this step, all segments are labeled with a class label, and a classified scale-sets structure is obtained. (4) Automatic optimal classification map selection: A multiscale accuracy assessment is conducted to evaluate the performances at different scales, and the optimal classification map is selected. A QuickBird image and a Gaofen-2 image were used to demonstrate the effectiveness and advantages of the proposed approach. The experimental results demonstrated that the optimal scale of an object-based image classification work is a range, not a single value. Moreover, the multiscale accuracies obtained using training samples and ground truth maps showed the same tendency. Therefore, it is possible to automatically estimate the optimal classification scale using training samples. Besides, we have demonstrated the effectiveness of using hierarchical enriching strategy to improve the performance of object-image classification.