MEcGANs

This repository contains data for the manuscript titled "Cloud Removal from Satellite Imagery using Multispectral Edge-filtered Conditional Generative Adversarial Networks". The implementation of the MEcGANs method presented in the manuscript is available at https://github.com/andrzejmizera/MEcGANs. The repository contains data that are based on freely available satellite images from the WorldView-2 European Cities dataset (https://earth.esa.int/eogateway/catalog/worldview-2-european-cities). Data provided by the European Space Agency. © ESA (2020). © DigitalGlobe, Inc. (2020), provided by European Space Imaging. Data.zip: - The 'images' folder contains target RGB, co-registered NIR, cloud-mask images, a real cloud RGB image and a co-registered real cloud NIR image used to synthesise clouded RGB and NIR images in accordance with the description provided in the manuscript. - The 'images/RGB' and 'images/NIR' images were obtained with the use of freely available satellite images from the WorldView-2 European Cities dataset. Images of various cities were selected and cropped into images of size 256x256. - The 'images/real_cloud_rgb_image.png' and 'images/real_cloud_nir_image.png' files contain an RGB image and a co-registered NIR image of a real cloud, respectively. They were obtained by extracting a part from one of large images available in the 'Landsat 8 Cloud Cover Assessment Validation Data' dataset (https://landsat.usgs.gov/landsat-8-cloud-cover-assessment-validation-data). - The 'datasets' folder contains lists of image filenames for the train and test subsets of the two datasets used in the experiments, i.e., the 'Berlin dataset' and the 'Paris dataset'. Results.zip: - Contains the complete sets of outputs generated by MEcGANs and McGANs for the two datasets used in the experiments for different values of NIR cloud penetrability, i.e., 1%, 0.5%, and 0.1%, as described in the manuscript. Jupyter notebooks for visual comparison of the generated images by the two methods, computation of the total difference contours area (TDCA) with respect to the corresponding ground truth images, and computation of four quality metrics, i.e., peak signal-to-noise ratio (PSNR), root mean squared error (RMSE), structural similarity index (SSIM), and spatial correlation coefficient (SCC), are provided.