Convolutional Neural Network-Driven Improvements in Global Cloud Detection for Landsat 8 and Transfer Learning on Sentinel-2 Imagery

Abstract

A stable and reliable cloud detection algorithm is an important step of optical satellite data preprocessing. Existing threshold methods are mostly based on classifying spectral features of isolated individual pixels and do not contain or incorporate the spatial information. This often leads to misclassifications of bright surfaces, such as human-made structures or snow/ice. Multi-temporal methods can alleviate this problem, but cloud-free images of the scene are difficult to obtain. To deal with this issue, we extended four deep-learning Convolutional Neural Network (CNN) models to improve the global cloud detection accuracy for Landsat imagery. The inputs are simplified as all discrete spectral channels from visible to short wave infrared wavelengths through radiometric calibration, and the United States Geological Survey (USGS) global Landsat 8 Biome cloud-cover assessment dataset is randomly divided for model training and validation independently. Experiments demonstrate that the cloud mask of the extended U-net model (i.e., UNmask) yields the best performance among all the models in estimating the cloud amounts (cloud amount difference, CAD = −0.35%) and capturing the cloud distributions (overall accuracy = 94.9%) for Landsat 8 imagery compared with the real validation masks; in particular, it runs fast and only takes about 41 ± 5.5 s for each scene. Our model can also actually detect broken and thin clouds over both dark and bright surfaces (e.g., urban and barren). Last, the UNmask model trained for Landsat 8 imagery is successfully applied in cloud detections for the Sentinel-2 imagery (overall accuracy = 90.1%) via transfer learning. These prove the great potential of our model in future applications such as remote sensing satellite data preprocessing.