Super-Scalable Computation Framework for Automated Terrain Identification

Automated labeling of very large-scale and high-resolution remote sensing imagery is an important application in machine learning and data science. We develop a parallel computation framework to accomplish pixel-level classification by fully utilizing the available resources. An effective superpixel tessellation is employed to group pixels into homogeneous regions and derive discriminating multi-pixel features. We deploy this segmentation method on multiple computing nodes instead of a single machine, thereby creating an efficient hierarchical representation for each image while drastically reducing performance bottlenecks. In the experiments, we further demonstrate that the identification accuracy of this approach is merely affected in a small range relative to that of a single machine, while the overall running time and handling capacity are significantly improved by using our super-scalable scheme.