Sampling based image splitting in large scale distributed computing of earth observation data

With increasing amounts of spatial, spectral and temporal remote sensing data and heterogeneity of platforms, we have entered an era of big data in remote sensing research. Imagery now routinely exceeds the memory size of personal computers so splitting/distributing big remote sensing data becomes a necessary pre-processing step. Standard rectangle based splitting methods can distort existing geometric and topological information and lose features as images are split into tiles. To address these challenges, we propose a sampling based image splitting method, which models the dataset as a streaming service and splits the dataset with a Voronoi diagram. The streaming data is systematically sampled to initially select the seeds of a Voronoi diagram. Voronoi regions are then generated according to spatial and spectral distances using Fortune's sweepline algorithm [1]. We test the splitting method with AVIRIS imagery of North America in 2013 (courtesy of NASA/JPL-Caltech) to evaluate the ability to detect objects of our splitting method. For evaluation we employ the object-based classification method of Hay and Castilla [2]. In contrast to rectangle based splitting approaches, most polygon borders generated by our method are found to converge with object borders (e.g., trees, building, and roads). When deployed with MapReduce, our sampling based splitting method also helps balance the computation intensity between each computing node.