The increasing availability of satellite information has improved Earth observation applications globally. However, primary satellite information is not as immediate as desirable. Indeed, besides the geometric and atmospheric limitations, clouds, cloud shadows, and haze generally contaminate optical imagery. Actually, such a contamination is intended as missing information and should be replaced. However, because the most common cloud masking algorithms take advantage by employing thermal images, here the objective is to provide an alternative algorithm suitable for multispectral imagery only. In addition, the work combines a multispectral/multitemporal approach as an effective method to retrieve daytime cloudless and shadow-free optical imagery. Experiment is undertaken upon mid- to low-spatial resolution data from Landsat 5 TM and Landsat 8 OLI, each for a different scene. A multitemporal stack, for the same image scene, is employed to retrieve a composite uncontaminated image over 1 year. The approach relies on a clouds and cloud shadows masking step, based on spectral features, a band-by-band multitemporal effect adjustment to avoid significant seasonal variations, and a data reconstruction phase based on automatic selection of the most suitable pixels from the stack. Results have been compared with a recognized masking algorithm approach and tested with uncontaminated image samples for the same scene. Accuracy and spectral features of the results provide high consistency.
The demand for remotely sensed data is growing increasingly, due to the possibility of managing information about huge geographic areas, in digital format, at different time periods, and suitable for analysis in GIS platforms. However, primary satellite information is not such immediate as desirable. Beside geometric and atmospheric limitations, clouds, cloud shadows, and haze generally contaminate optical images. In terms of land cover, such a contamination is intended as missing information and should be replaced. Generally, image reconstruction is classified according to three main approaches, i.e. in-painting-based, multispectral-based, and multitemporal-based methods. This work relies on a multitemporal-based approach to retrieve uncontaminated pixels for an image scene. We explore an automatic method for quickly getting daytime cloudless and shadow-free image at moderate spatial resolution for large geographical areas. The process expects two main steps: a multitemporal effect adjustment to avoid significant seasonal variations, and a data reconstruction phase, based on automatic selection of uncontaminated pixels from an image stack. The result is a composite image based on middle values of the stack, over a year. The assumption is that, for specific purposes, land cover changes at a coarse scale are not significant over relatively short time periods. Because it is largely recognized that satellite imagery along tropical areas are generally strongly affected by clouds, the methodology is tested for the case study of the Dominican Republic at the year 2015; while Landsat 8 imagery are employed to test the approach.