Segmentation of large-scale remotely sensed images on a Spark platform: A strategy for handling massive image tiles with the MapReduce model

Abstract

Image segmentation is essential in object-based image analysis. Numerous image segmentation algorithms have been proposed and widely applied to process remote sensing images, but most of them are designed to deal with single scenes. As the volume of images grows rapidly, handling images with single machines is becoming increasingly difficult, and the size of a composite image can be larger than the CPU memory of a single computer. To address this problem, a distributed image segmentation strategy is proposed in this paper. The two main steps of the proposed strategy are as follows. First, a prepared massive image is loaded and then decomposed into sub-images that are distributed across multiple computers; algorithms are then used in parallel to segment each sub-image into a large number of initial objects. Secondly, the proposed object resegmentation method is applied to the initial boundary objects in each sub-image in order to merge these objects. The sub-images are then ingested from the different computers in order to obtain the final segmentation image. Two classical segmentation algorithms are employed to test the proposed strategy in eight different study areas that include urban area, suburban zone and agricultural landscape. Both the intersection over union and the F-measure metrics show that the proposed strategy can help to solve the problem of the data volume being too large to fit on a single machine, and that it also performs better than comparative strategies. The proposed strategy not only has the ability to segment very large images, but also accelerates the segmentation and segmentation-based applications so that they can match the image acquisition rate.