A CNN cloud detector for panchromatic satellite images

Cloud detection is a crucial step for automatic satellite image analysis. Some cloud detection methods exploit specially designed spectral bands, other base the detection on time series, or on the inter-band delay in push-broom satellites. Nevertheless many use cases occur where these methods do not apply. This paper describes a convolutional neural network for cloud detection in panchromatic and single-frame images. Only a per-image annotation is required, indicating which images contain clouds and which are cloud-free. Our experiments show that, in spite of using less information, the proposed method produces competitive results.