An Assessment of Geometric Activity Features for Per-pixel Classification of Urban Man-made Objects using Very High Resolution Satellite Imagery

摘要

In this paper, we propose the use of Geometric Activity (GA) features for detecting man-made objects in urban areas using VHR satellite imagery. These features describe the geometric context of a pixel without the necessity of segmentation and can beintegrated as extra bands in a per-pixel classification. Two main types of GA features were investigated: ridge features based on the well-known facet model and morphological features obtained by applying closing transforms with structuring elements of different size and shape. Our findings show a substantial increase in classification accuracy for the man-made object classes "roads and buildings with dark roof" after inclusion of GA features. Next to GA features, the use of object-based features derived from eCognition~R, containing both geometric and textural information, was also investigated for per-pixel classification. Accuracies obtained with object-based features are comparable to the accuracies obtained with GA features. The inclusion of bothGA features and object-based features further improves the overall accuracy. GA features and object-based features thus contain complementary information.