Comparison of histogram-curve fitting-based and global threshold methods for cloud detection

The goal of this research is fit the best curve to the image histogram using polynomial (i.e., linear, quadratic, cubic, and quartic), Fourier and Gaussian with two different term numbers, and Bezier functions, and then compare their performance with nine automatic state-of-the-art thresholding methods: Shanbhag, Triangle, Huang, Intermode, IsoData, Li's, Maximum entropy, Moment, and Otsu. The visible bands of the Landsat-9, Landsat-8, and Sentinel 2 satellite images are averaged to create an index map, which was used to perform the cloud detection. Since many satellites system only has visible bands, the average index map was used. Three different images which include various cloud types were clipped from each satellite imaging. Thresholding methods were evaluated using statistical tools such as structural similarity index measure (SSIM), peak signal-to-noise ratio (PSNR), and universal image quality index (UIQI), accuracy, Cohen's kappa, precision, recall, F-Measure, and G-Mean. The histogram-based approaches that performed the best after Li's method were Bezier (96.71% accuracy and 99.83% SSIM), 2nd Gaussian (97.62% precision), Quartic (80.56% kappa, 16.16% PSNR, 82.45% F-Measure, and 89.52%), and 2nd Fourier (81.08% recall and 89.27% G-Mean), according to the average statistics of all images. Additionally, methods other than Shanbhag yielded favorable results (more than 90% accuracy).