Author
 |
MUELLER WARRANT, GEORGE |
|
Submitted to: International Journal of Applied Earth Observation and Geoinformation
Publication Type: Peer Reviewed Journal Publication Acceptance Date: 2/12/2019 Publication Date: 3/8/2019 Citation: Mueller Warrant, G.W. 2019. Multistep block mapping on principal component uniformity to repair Landsat 7 defects. International Journal of Applied Earth Observation and Geoinformation. 79:12-23. https://doi.org/10.1016/j.jag.2019.02.005. DOI: https://doi.org/10.1016/j.jag.2019.02.005 Interpretive Summary: Remote sensing classification is a powerful tool for tracking crop patterns across landscapes, but defect gaps in imagery either from failure of the Landsat 7 scan line corrector and/or the presence of scattered clouds severely limit the number of high quality images available for analysis in many locations. A robust method for estimation of missing data in imagery with defects was developed by researchers in Corvallis, Oregon using Python scripts and ArcGIS tools. Unlike most other methods for filling in the gaps, this new procedure includes statistical information on the reliability of the individual estimates, allowing users to fine-tune their selection of how much of the repaired data to use. Although it was developed for the complex landscapes and cropping patterns of the Pacific Northwest, our testing validated the reliability of its performance in a variety of other settings, including the upper Midwest of the United States, subtropical southeastern China, and the Mexican desert. The improved quality and increased numbers of images this procedure now makes available is anticipated to increase the number of landuse categories able to be detected, improve overall classification accuracy, and simplify steps needed to produce final versions of landuse classification and multiyear crop rotation history. Technical Abstract: Permanent failure of the Landsat 7 Scan Line Corrector (SLC) in May 2003 left 22% of scene areas unscanned, seriously impairing the scientific utility of all subsequently collected imagery and leading to the development of a variety of schemes to produce reasonable estimates for missing data. All existing procedures to restore data tend to possess one or more of the following limitations: (1) the need to arbitrarily set parameters such as numbers of classes into which the landscape may be subdivided in later analyses, (2) residual visual striping and other artifacts, (3) inadequate knowledge of the statistical properties of errors associated with estimates for missing pixels, and (4) challenges in scaling procedures up to the size and scope of the entire post SLC-failure Landsat 7 archive. We developed a robust procedure for filling in missing data from means of pixels present within 2**N–sized neighboring blocks defined by simultaneously minimizing heterogeneity of source imagery within blocks while also satisfying the need for large enough blocks that not all the pixels within them were missing. Selection of regions used to fill in the missing data involved a stepwise process in which local standard error cutoffs for acceptance of any particular 2**N–sized region of source imagery as best representing missing pixels potentially located within it were progressively raised over a series of 50 cycles, followed by one final step, when necessary, to fill in remaining gaps. |
