Mitigation of Weather effects on Optical Satellite Imagery

Abstract

How to deal with the presence of weather affected data is an unavoidable topic in the processing of optical imagery. Clouds and cloud shadows significantly alter the spectral signatures obtained from satellite data, which often leads to problems for any kind of scientific analysis. In this research there has been elaborated on two different kind of problems: The detection of clouds and cloud shadows and the mitigation of the effect caused by cloud shadows. Most of existing operational cloud detection algorithms are so-called rule-based. Their performance is highly variable and they have their limitations. A new promising research was done by Mohajerani and Parvaneh (2019), where a convolutional neural network (CNN) named ’Cloud-Net’ was developed. In this study we have elaborated on this CNN, by converting the analysis to Sentinel-2 data and making significant modifications on the model setup. The results have been compared to the ESA Scene Classification Map (SEN2COR algorithm). It was found that for the detection of clouds the overall CNN accuracy outperforms the ESA Scene Map (95.6% vs. 92.0% respectively). For the detection of cloud shadows the modified Cloud-Net model also gave better results (90.4% vs. 84.4%). Previous work on cloud shadow correction algorithms show rather complex and inconvenient methods, where the only goal was to remove the effect of the shadow. If one is interested to also correct for illumination effects, to make it more aligned to a predetermined ground truth, new possibilities arise which allows for simpler and more direct methods. Two proposed methods have been investigated in this study. The first method, called ’decomposition of components’, investigated the use of a single formula. The affected cloud shadow pixel is corrected based on the RGB difference with a ground truth image, and a single correction factor that was determined based under the assumption that cloud shadows cause a homogeneous alteration effect in a small area. The second method, called the ’CNN based method’, presents a totally new idea by changing the Cloud-Net model to a regression model, in order to correctly alter cloud shadow affected pixels. The performance of both methods was quantified by the structural similarity index measure (SSIM). It was found that the decomposition of components method has the most potential, showing significant improvements on the correction of cloud shadow affected areas.