Multi-scale RoIs selection for classifying multi-spectral images

Abstract

The applications of object-based image analysis (OBIA) in remote sensing studies have received a considerable amount of attention over the recent decade due to dramatically increasing of the spatial resolution of satellite imaging sensors for earth observation. In this study, an unsupervised methodology based on OBIA paradigm for the estimation of multi-scale training sets for land cover classification is proposed. The proposed method consists of selection of valid region of interests in an unsupervised way and its characterization using some attributes in order to form meaningful and reliable training sets for supervised classification of different land covers of a satellite image. Multi-scale image segmentation is a prerequisite step for estimation of multi-scale training sets. However, scale selection remains a challenge in multi-scale segmentation. In this work, we propose a method to determine the appropriate segmentation scale for each land cover with the help of prior knowledge in the form of in-situ data. The proposed method is further discussed and validated through multi-scale segmentation using quick shift and random forest algorithms on two multi-spectral images captured using Worldview-2 sensor. Experimental results indicate that the proposed method qualitatively and quantitatively outperforms three state-of-the-art methods.