Improvement on the three-step haze removal technique with the aid of one clear image partly overlapped

In our previous studies, we have proposed a three-step haze removal technique (Haze detection, Haze perfection, Haze removal) to solve spatial varying haze contamination in multispectral satellite imagery, which was originally developed for individual hazy images. Haze detection step results in a Haze Thickness Index (HTI) image from the hazy image, Haze perfection step corrects spurious value in HTI image, and Haze removal step dehazes the hazy image with the aid of HTI image. In the Haze removal step, the best method is Virtual Cloud Point (VCP) with the flaw of too much human intervention in parameter determination to obtain a suitable VCP, which makes this technique slightly user dependent. In this study, we improve the three-step haze removal technique by automating VCP method with the aid of one clear image partly overlapped. Case data are two partly overlapped QuickBird images, one clear and one hazy with five-day interval. After Haze detection and Haze perfection steps, we delineate 76 paired (the same object in hazy and clear images) polygon samples in the overlapped region of the two images respectively, and obtain the mean HTI and digital number (DN) of each sample. Consequentially, we will obtain 76 dehazed samples if we implement Haze removal step on the 76 hazy samples, and we are able to assess the efficiency by calculating the correlation coefficient between 76 paired dehazed DN and clear DN. The improvement in this study is optimizing the parameters in VCP method by the Hooke-Jeeves algorithm (also named as Pattern Search Method), in order to carry out the maximum correlation coefficient. Hooke-Jeeves algorithm requires three parameters determined before start: starting VCP (two dimensional vector [HTIvcp, DNvcp]), step, and minimum step. To a continuous function, the three parameters are only responsible for iterative times, but not the final result. In our study, since the formula contained in VCP method is not continuous, HTIvcp of the staring VCP must be larger than the maximum HTI of the 76 samples. The result shows that, the largest correlation coefficients of four bands are 0.886608, 0.873472, 0.909047, 0.936123 respectively after improvement, which are a little better than the result from VCP by human intervention (0.847567, 0.838695, 0.889095, 0.904616). Though the improvement doesn't take a big step forwards, it makes the three-step haze removal technique more objective.