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

Authors: Mueller Warrant, George; Whittaker, Gerald

Submitted to: International Journal of Remote Sensing
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 9/26/2015
Publication Date: 11/16/2015
Citation: Mueller Warrant, G.W., Whittaker, G.W. 2015. Distorted spatial warping to compress large, high resolution rasters for remote sensing classification. International Journal of Remote Sensing. 36(22):5664-5689.

Interpretive Summary: Decisions made early in the data preparation phases of a remote sensing classification project set fundamental limits on the value to society of the final product. The two most commonly used methods for incorporating images with widely varying resolutions both have serious although complementary limitations. Degrading high resolution imagery such as National Agriculture Imagery Program (NAIP) to match lower resolution data such as Landsat in essence discards over 99% of the information present at high resolution with no knowledge of how valuable those lost data might be. Current methods for enhancing the resolution of imagery such as Landsat are generally limited to approximately four-fold increases in data density, a far cry from the 900-fold difference commonly present between Landsat and high resolution satellite data or NAIP. Even if methods were developed that could convert Landsat data to more closely match the density of high resolution data, the resulting rasters would likely be too large to use on most computers commonly conducting remote sensing classification projects. Our new method for aligning data across wide ranges in resolution measures the variability present in the high resolution data over a variety of spatial scales called neighborhoods and uses this information to decide how strongly to compress the data in each possible neighborhood over the entire image. The compressed images are highly warped, include multiple neighborhood sizes, and not suited for direct visual interpretation. They can be used in existing remote sensing classification software and were found to generate significant improvements in accuracy even when the high resolution data were compressed to nearly the same average level used in the typical uniform degradation involved in converting from 1 m on up to 30 m resolution. Because of the growing availability of data with widely varying resolutions, we believe our method has a good chance of adoption by other users as a behind-the-scenes component to improvements in accuracy, number of categories included in classifications, and geographic extent of analyses. Patent potential is being actively pursued along with searches for software development partners.

Technical Abstract: The often used approach of degrading/downsampling high resolution (~ 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, but does so by forgoing all ability to detect features smaller than 30 m in addition to discarding over 99% of the information content of the high resolution data. The alternative of upsampling coarser resolution data into smaller-sized synthetic pixels creates its own set of problems, including enormous file sizes, likely absence of meaningful variation over small spatial scales, and no assurance of meaningful improvement in classification accuracy despite the guaranteed increase in computational time and resources. We propose a new method for compressing data and generating aligned, multiband rasters needed for remote sensing classification by using local variability in the finest resolution data available to define areas adequately represented by mean values averaged over neighborhood sizes of increasing powers of 3, with neighborhoods ranging from a maximum of 81 by 81 pixels on down to a minimum of the original high resolution pixels. We achieved compression of 138-fold in disk storage and 785-fold in actual counts of non-null pixels through our choice of cut-off values for accepting 3 by 3, 9 by 9, 27 by 27, or 81 by 81 regions of tolerable variability, while otherwise retaining full resolution data in regions 3 cells wide by 3 cells high. When applied to a remote sensing classification of the Willamette Valley of Oregon, 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 our 55 existing ground-truth classes, leading to further increases in accuracy.