Relative gain characterization and correction for pushbroom sensors based on lifetime image statistics and wavelet filtering

The primary objective of this thesis is to develop a tool that is able to reduce stripes from imagery acquired by pushbroom sensors caused by detector-to-detector non-uniform response. Image uniformity in the pushbroom sensor requires uniform detector response across an array of detectors. If the response of the detector is not uniform, stripes and other artifacts may occur in the image. This work primarily concentrates on two methods to minimize stripes from an image. The first method estimates the best set of relative gains by identifying the best type of image to use for this estimate based on image mean and standard deviation. The second method is based on cosmetic de-striping that filters most of the stripes from an image. The former method implements the technique of histogram equalization to estimate relative gains from the lifetime image statistics data. This method provides two estimates of relative gain; one based on the mean and the other based on the standard deviation. The estimated relative gain can then be applied to the image which reduces stripes from the image. The other method that deals with minimization of stripes using a cosmetic de-striping approach is based on wavelet filtering. Three different algorithms based on wavelets are derived: Low Frequency Sub-band (LFSB), High Frequency Sub-band (HFSB) and All Frequency Sub-band (AFSB). In each of these approaches, the main idea is to decompose the image into different frequency components using a wavelet transform, apply an appropriate filter to various image components to remove stripes, and reconstruct the image using a corresponding inverse wavelet transform. In order to validate these techniques, these algorithms were implemented on images acquired by the ALI sensor carried by the EO-1 satellite. The results were analyzed qualitatively and quantitatively. The analysis suggested that images with high mean and high standard deviation (HMHSD) are best to estimate the relative gains, because relative gain estimates from these scenes results in less stripes after correction and require less scene to stabilize the relative gain. When preference was given to HMHSD scenes, the estimates of relative gain based on mean performed well compared to the estimates of relative gain based on standard deviation. Similarly, the HFSB approach with two levels of sub-banding was observed to be sufficient in minimizing stripes and retaining most of the information of the image.