The famous blue marble images so often used on maps and globes were stitched together from data produced by the Moderate Resolution Imaging Spectroradiometer (MODIS) satellite imager instrument, which can be found on NASA’s Aqua and Terra Satellites. Because MODIS images are high quality, free, and easily available, they have been used in a wide range of scientific research. Unlike a conventional camera with R, G, and B bands, MODIS has 36 bands covering the visible as well as the near, shortwave and thermal infra-red spectrum. In particular, the shortwave infra-red 1.6-μm band 6 is critical for distinguishing ice and snow from clouds. This is because, while all three substances are highly reflective, ice and snow are significantly darker than clouds when viewed at 1.6-μm. Unfortunately instruments are often damaged by the harsh conditions in space. In the case, of MODIS on Aqua, 15 out of the 20 detectors that produce the band 6 image are damaged which means that ¾ of the scan-lines are effectively missing.

In this work we use the other bands, as well as the 5 working detectors, to estimate the full band 6 image. Even though a values in band 6 is not completely determined by the values in the other bands at that pixel, it can be well approximated as a function of a small (typically 5x5) local window across the other bands; thus local texture plays a role in the statistical relationship. So in a local tile (larger than our local window), we can use the 5 working detectors to fit a linear estimator from the other MODIS bands and then apply that estimator to restore the data missing from the remaining 15 detectors. Fitting the coefficients globally for local window across the image does not work well because the regression coefficients change gradually within the image. Instead we when we restrict our regression to large 200x200 overlapping tiles and smoothly transition across overlaps with a smooth blending of the regression coefficients.

Because we obtained promising results in the paper describing the algorithm, for the radiance and reflenctance values we tested the use of the restored band 6 in subsequent papers for cloud and snow products. For example, we evaluated by restoring a band 6 with simulated damage and used the MODIS/Terra snow algorithm to produce a snow product and then compared with a snow product using the actual band 6. This gave nearly identical results. We established this both for cloud and snow products. Given the critical role of band 6 in the snow product, we collaborated with the group at NASA responsible for the snow product, Dorothy Hall and George Riggs. and as a result our restoration is now part of the latest release of the NASA product, collection 6. Hall even referred to the restoration of band 6 as a "game changer."