An Object-Oriented Approach to the Classification of Roofing Materials Using Very High-Resolution Satellite Stereo-Pairs

Abstract

The availability of multispectral images, with both high spatial and spectral resolution, makes it possible to obtain valuable information about complex urban environment, reducing the need for more expensive surveying techniques. Here, a methodology is tested for the semi-automatic extraction of buildings and the mapping of the main roofing materials over a urban area of approximately 100 km2, including the entire city of Bologna (Italy). The methodology follows an object-oriented approach and exploits a limited number of training samples. After a validation based on field inspections and close-range photos acquired by a drone, the final map achieved an overall accuracy of 94% (producer accuracy 79%) regarding the building extraction and of 91% for the classification of the roofing materials. The proposed approach proved to be flexible enough to catch the strong variability of the urban texture in different districts and can be easily reproducible in other contexts, as only satellite imagery is required for the mapping.

Keywords

• building extraction
• roofing classification
• WorldView-3
• OBIA classification
• machine learning
• smart city