On a Structural Similarity Index Approach for Floating-Point Data.

Area of Science:

• Scientific Computing
• Data Analysis
• Climate Modeling

Background:

• Data visualization is crucial for analyzing large simulation outputs, often involving image generation for evaluation.
• The Structural Similarity Index Measure (SSIM) is commonly used for image comparisons, but creating numerous images from large datasets is computationally expensive.

Purpose of the Study:

• To introduce a novel metric, the Data SSIM (DSSIM), as a computationally efficient alternative to SSIM for assessing data quality.
• To evaluate the DSSIM's effectiveness in quantifying differences in large-scale climate model simulation data, particularly after lossy compression.

Main Methods:

• Developed the Data SSIM (DSSIM) metric, which operates directly on floating-point simulation data.
• Applied the DSSIM to quantify data differences resulting from lossy compression in a climate model dataset.
• Compared the performance and utility of DSSIM against traditional SSIM-based image comparison methods.

Main Results:

• The DSSIM significantly improves performance by eliminating the need for intermediate image creation.
• DSSIM effectively quantifies data quality differences, demonstrating its utility in evaluating lossy compression effects.
• The DSSIM avoids data-independent, plot-specific choices that can influence SSIM results, leading to more robust data quality assessments.

Conclusions:

• The DSSIM offers a computationally efficient and robust method for assessing data quality in large simulation datasets.
• This metric is particularly valuable for workflows involving large volumes of floating-point data, such as climate modeling.
• The DSSIM has the potential for broader application in scientific computing and data analysis beyond climate science.