LARGE-SCALE IMAGE PROCESSING USING MASS PARALLELIZATION TECHNIQUES

Assets of raw geo-located imagery can be divided into tiles and coverage masks can be generated for each tile. For each tile, fragments of pixels from coverage masks of neighboring tiles can be extracted and tagged. The fragments can be sorted and stored in a data structure so that fragments having the same tag can be grouped together in the data structure. The fragments can be used to feather the coverage mask of the tile to produce a blend mask. Multi-resolution imagery and mask pyramids can be generated by extracting fragments from tiles and minified (e.g., down-sampled). The minified fragments can be tagged (e.g., by ancestor tile name), sorted and stored in a data structure, so that fragments having like tags can be stored together in the data structure. The fragments can be assembled into fully minified tiles for each level in the pyramid. Input tiles in a first projection are re-projected into a second projection using techniques that minimize distortion in the re-projected imagery.