A Statistical Approach to Obtaining a Data Structural Similarity Index Cutoff Threshold

Huge climate simulations such as the Community Earth System Model Large Ensemble (CESM-LE) output massive amounts of data, and switching from lossless to lossy compression of this data is inevitable. Applying lossy compression to climate data requires a guarantee that scientific conclusions will not be affected by the compression process. One way to check if a scientist's conclusions will be altered is by assessing if the data has been visually altered in any way. To that extent, the Data Structural Similarity Index Measure (DSSIM) is designed to test the visual similarity between two images. It is based on the well-known Structural Similarity Index Measure (SSIM), but has been modified to fit our application. The DSSIM is computed on two datasets instead of two images. This makes the measure invariant to the plotting parameters. Additionally, the DSSIM is much more efficient to compute on a floating-point dataset than the SSIM. A user study in a previous work determined an appropriate SSIM threshold, above which images are highly likely to be visually indistinguishable. In this work, we use the results of the previous study and statistical techniques to translate the SSIM threshold to an appropriate cutoff threshold for the DSSIM. We find the appropriate threshold by minimizing the pass/fail classification difference between images classified using the DSSIM and the previously acquired SSIM threshold. This threshold results in agreement between the image classification results using either metric in more than 92% of cases.