Alexander L. De Souza , Parnia Shokri

Semantic segmentation of clouds in Earth observation imagery is an important task in a variety of remote sensing contexts: from the application of atmospheric corrections to being able to accurately omit cloud pixels when extracting information about ground features. Here we introduce a deep learning approach based on the popular U-Net architecture. The core of the architecture is an U-Net with residual units that ease the training of the network. An attention mechanism is also incorporated to enable the model to more effectively learn and distinguish between cloud and non-cloud features. We also explore two complementary loss functions, Binary Cross Entropy and Jaccard, in order to overcome data imbalances common to this application. Our model is trained on a uniquely curated dataset spanning a wide variety of resolutions, scene contexts, lighting conditions, and seasonality. Our experiments demonstrate that this model is an accurate and robust model for the semantic segmentation of clouds in satellite imagery, and the model achieves state-of-the-art performance over many other models (including others based on CNN architectures) on common benchmark datasets, even without having been exclusively trained on images from the sources in those datasets.