The challenges of image segmentation in big remotely sensed imagery data

Abstract

With the increase in spatial, spectral and temporal resolutions of Earth observing systems, geospatial and remote sensing (RS) image research is shifting towards a big data paradigm. One of the most important challenges in RS big data is image segmentation, which is defined as a process to group pixels together by a predefined criteria. Image segmentation allows for the extraction of features such as roads or habitats or buildings. Image segmentation is rendered more difficult with big data because the computing power on single platforms cannot keep pace with the size and velocity of new data. Big data sets must be decomposed for the analysis in distributed and parallel computing platforms. Decomposition through techniques like slicing by spatial extent obscures the geometric and topological information in geospatial data, for example generating fake artefacts. To address these challenges, we propose a geospatial cyberinfrastructure (GCI) that coordinates cloud computing, MapReduce framework, image segmentation algorithms, a spatial extent splitting method and a recomposing technique using moving window. This GCI is evaluated on cloud computing to identify features in a 312.07 GB high-resolution colour aerial photo with Hadoop. K-means-based image segmentation is selected as the case study. We deploy the architecture in private cloud and public cloud implementation, respectively. The results demonstrate the benefits of the decomposing and recomposing methods in segmenting images, removing fake artefacts and reducing information distortion. More general problems in big data are revealed, among them I/O problems, particularly in the amount of preprocessing and post-processing that will be required in any analysis of big imagery data. We conclude with implications for scalability and suggestions to speed up decomposition and recomposition.