A new method for generating a clear-sky Landsat composite for cropland from cloud-contaminated Landsat-7 and Landsat-8 images

A new method was developed in this study for producing a clear-sky Landsat composite for cropland from cloud-contaminated Landsat images acquired in a short time period. It used Thiel–Sen regression to normalize all Landsat scenes to a MODIS image to make all Landsat images radiometrically consistent and comparable. Pixel selection criteria combining the modified maximum vegetation index and the modified minimum visible reflectance selection methods were designed to enhance the pixel selection of land/water over cloud/shadow in the image compositing. The advantages of the method include (1) avoiding complicated atmospheric corrections but with reliable surface reflectance results, (2) being insensitive to errors induced by image co-registration uncertainties between Landsat and MODIS images, (3) avoiding the lack of samples for the regression analysis using the full Landsat scenes (rather than overlay regions), and (4) enhancing cloud/shadow detection. The composite image has MODIS-like surface reflectance, thus making MODIS algorithms applicable for retrieving biophysical parameters. The method was automatically implemented on a set of 13 cloud-contaminated (>39%) Landsat-7 (Scan-Line Corrector-Off) and Landsat-8 scenes acquired during peak growing season in a crop region of Manitoba, Canada. The result was a 95.8% cloud-free image. The method can also substantially increase the usage of cloud-contaminated Landsat data.