Detekce oblačnosti z družicových dat pomocí neuronových sítí

Abstract

The presence of cloud cover is one of the most significant obstacles in processing optical satellite data. Traditional detection methods, forming part of official products in missions such as Sentinel-2, are based on spectral threshold values and physical rules. However, these fixed rules fail in areas where clouds spectrally overlap with other surfaces – for example snow, bright soils, or water bodies. This work focuses on the use of deep learning for cloud detection in Sentinel-2 data. Four segmentation models were trained, combining U-Net and DeepLabV3+ architectures with ResNet34 and MobileNetV2 backbone networks, on the CloudSEN12+ dataset (8,490 training, 535 validation, and 975 test patches). The models were evaluated on the test portion of CloudSEN12+, on 75 custom Sentinel-2 images, and – to verify transferability – on Landsat 8 SPARCS data. On Sentinel-2 data, the models achieved F1 scores ranging from 82-95% depending on the location and model configuration, outperforming traditional algorithms in most cases. In the overall comparison, the OmniCloudMask model performed best with an F1-score of 90-98% depending on location, while the QA60 mask showed the weakest results with F1-scores between 11.2 and 67%. Although augmentation techniques were used during model training to simulate varying spatial resolutions, a drop in F1-score was observed when testing on SPARCS data (30 m resolution). The ResNet34 variants achieved F1-scores of around 50%, while the MobileNetV2 variants performed at approximately 40%. The best results among all tested methods were provided by the OmniCloudMask model, with an F1-score of 88%.