Distorted spatial warping to compress large, high-resolution rasters for remote-sensing classification

Abstract

Decisions made early in the data preparation phases of remote-sensing classification projects set fundamental limits on the value to society of the final products. The often-used approach of degrading/down-sampling high-resolution (e.g. 1 m pixel size) imagery to match lower-resolution data (e.g. Landsat 30 m) through averaging or majority-rule solves the problem of aligning pixels across bands of differing resolution, but does so by forgoing all ability to detect features smaller than 30 m in addition to potentially discarding up to 99% of the information content of the high-resolution data. The alternative of up-sampling coarser-resolution data into smaller-sized synthetic pixels creates its own set of problems, including potentially enormous file sizes, likely absence of meaningful variation over small spatial scales (which may generate matrix singularities fatal to the maximum likelihood classifier), and no assurance of meaningful improvement in classification accuracy despite guaranteed increases in computational time and resource requirements. We propose a new ‘warped space compression technique’ as a variation of vector quantization that analyses local variability in the finest-resolution data available to define acceptable pixel-based neighbourhood (N × N) sizes over which data can be averaged while minimizing overall information loss. Alternative neighbourhoods are aligned so that nine smaller ones nest within each progressively larger one as 3 × 3 squares, resulting in local data compression options of 3 × 3 (ninefold), 9 × 9 (81-fold), 27 × 27 (729-fold), and 81 × 81 (6561-fold). Our transformation process to ‘warped space’ created spatially distorted images with jagged east edges and little visually discernible relationship to the original data. We achieved compressions of 48- to 138-fold in disc storage and 292- to 785-fold in actual numbers of non-null pixels through our choice of cut-off values for accepting 3 × 3, 9 × 9, 27 × 27, or 81 × 81 neighbourhoods of tolerable variability, while otherwise retaining full (1 m) resolution data in regions three cells wide by three cells high. Medium-resolution data (e.g. Landsat 30 m) can be translated into the warped space defined by high-resolution data and composited with it for conducting remote-sensing classifications. When applied to a 71-band, 55-class remote-sensing classification of a 25,500 km2 region centred on the Willamette Valley of Oregon, USA, classification accuracy increased from 64.4% in normal space to 71.3% in warped space. Unsupervised classification in warped space identified several additional categories that could be appended to the 55 existing ground-truth classes, leading to further increases in accuracy. Warped-space compression may be particularly beneficial for ecological studies where it could maintain high resolution in features of interest such as riparian buffers without creating exorbitantly large data files.