Krupiński M., Lewiński S., Malinowski R. (2019), One class SVM for building detection on Sentinel-2 images, Proc. SPIE
Human population has reached 7.7 billion in year 2019. According to the forecasts of world population projections it may reach 8 billion in 2023. This growth indicates inter alia increased demand for housing. Monitoring of urbanization processes in large scale may be effectively monitored only using Earth Observations (EO).
European initiative – the Copernicus program provides unique opportunity for urban areas monitoring with constellation of EO satellites. All data gathered by these satellites are available for any user with no extra cost. Sentinel-2 (S2) satellites provide the highest spatial resolution among satellite of Sentinels constellation. Depending on spectral band it varies between 10 m, 20 m and 60 m. 10 m and 20 m pixels are detailed enough to recognize bigger building structures and have been used in this research.
Satellite image classification in most cases is conducted as a multi-class analysis. Each pixel is assigned to one of the number of land cover classes according to specific algorithm rules. This approach requires knowledge about all the classes occurring on the image. However, in case when information only about one class is available, specific type of algorithms may be applied, the so called one class classifier (OCC). In that case, knowledge about one land cover class should be sufficient to perform classification task. One class Support Vector Machine (OCSVM) is a state-of-art OCC algorithm which has been used mainly for very high resolution satellite imagery and aerial image classification. Li at al. has used OCSVM for building damage detection, urban road damage detection and four other classes separately (urban, trees, grass and soil) by Li et al. Aerial hyperspectral image has been classified with OCSVM to identify tree species by Baldeck et al. The same classifier has been also used for time series analyses to map raised bogs on very high resolution satellite images and paddy rice on low resolution satellite images (MODIS). Paddy rice was also classified with OCSVM on high resolution image (Landsat 8) by Xu et al. Mack et al.11 tested OCSVM on layer stacks made from three optical images and three radar images. According to the performed literature review, OCSVM has not been analyzed in the context of Sentinel-2 images classification.
In our research, the OCSVM classifier has been used for the first time with high resolution satellite images from Sentinel-2, for the task of buildings detection. We analyzed different algorithm parameters and proposed a visualization tool, which facilitates selection of the optimum values for them. Moreover, we evaluated automatic solution for selection and tuning of training data. The last issue concerned the analysis of variability of accuracy depending on the time of image acquisition.
Krupiński M., Lewiński S., Malinowski R. (2019), One class SVM for building detection on Sentinel-2 images, Proc. SPIE 11176, Photonics Applications in Astronomy, Communications, Industry, and High-Energy Physics Experiments 2019, 1117635 (6 November 2019); https://doi.org/10.1117/12.2535547