Optimizing WorldView-2, -3 cloud masking using machine learning approaches

The detection of clouds is one of the first steps in the pre-processing of remotely sensed data. At coarse spatial resolution (> 100 m), clouds are bright and generally distinguishable from other landscape surfaces. At very high-resolution (< 3 m), detecting clouds becomes a significant challenge due to the presence of smaller features, with spectral characteristics similar to other land cover types, and thin (partially transparent) cloud forms. Furthermore, at this resolution, clouds can cover many thousands of pixels, making both the center and boundaries of the clouds prone to pixel contamination and variations in the spectral intensity. Techniques that rely solely on the spectral information of clouds underperform in these situations. In this study, we propose a multi-regional and multi-sensor deep learning approach for the detection of clouds in very high-resolution WorldView satellite imagery. A modified UNet-like convolutional neural network (CNN) was used for the task of semantic segmentation in the regions of Vietnam, Senegal, and Ethiopia strictly using RGB + NIR spectral bands. In addition, we demonstrate the superiority of CNNs cloud predicted mapping accuracy of 81–91%, over traditional methods such as Random Forest algorithms of 57–88%. The best performing UNet model has an overall accuracy of 95% in all regions, while the Random Forest has an overall accuracy of 89%. We conclude with promising future research directions of the proposed methods for a global cloud cover implementation. •The CNN models were able to generalize consistently through both space and time.•The CNN model has a superior overall accuracy of 94.9%, 6.4% higher than the RF.•The CNN cloud model is robust to seasonality effects over a 12-year period.•Further investigation is required to assess effectiveness in mid/high latitudes.•The CNN models are robust enough to leverage diversified location-time training.