Cloud masking is of central importance to the Earth Observation community. This paper deals with the problem of detecting clouds in visible and multispectral imagery from high-resolution satellite cameras. Recently, Machine Learning has offered promising solutions to the problem of cloud masking, allowing for more flexibility than traditional thresholding techniques, which are restricted to instruments with the requisite spectral bands. However, few studies use multi-scale features (as in, a combination of pixel-level and spatial) whilst also offering compelling experimental evidence for real-world performance. Therefore, we introduce CloudFCN, based on a Fully Convolutional Network architecture, known as U-net, which has become a standard Deep Learning approach to image segmentation. It fuses the shallowest and deepest layers of the network, thus routing low-level visible content to its deepest layers. We offer an extensive range of experiments on this, including data from two high-resolution sensors—Carbonite-2 and Landsat 8—and several complementary tests. Owing to a variety of performance-enhancing design choices and training techniques, it exhibits state-of-the-art performance where comparable to other methods, high speed, and robustness to many different terrains and sensor types.

No abstract

We introduce a novel neural network architecture-Spectral ENcoder for SEnsor Independence (SEnSeI)-by which several multispectral instruments, each with different combinations of spectral bands, can be used to train a generalised deep learning model. We focus on the problem of cloud masking, using several pre-existing datasets, and a new, freely available dataset for Sentinel-2. Our model is shown to achieve state-of-the-art performance on the satellites it was trained on (Sentinel-2 and Landsat 8), and is able to extrapolate to sensors it has not seen during training such as Landsat 7, Per úSat-1, and Sentinel-3 SLSTR. Model performance is shown to improve when multiple satellites are used in training, approaching or surpassing the performance of specialised, single-sensor models. This work is motivated by the fact that the remote sensing community has access to data taken with a hugely variety of sensors. This has inevitably led to labelling efforts being undertaken separately for different sensors, which limits the performance of deep learning models, given their need for huge training sets to perform optimally. Sensor independence can enable deep learning models to utilise multiple datasets for training simultaneously, boosting performance and making them much more widely applicable. This may lead to deep learning approaches being used more frequently for on-board applications and in ground segment data processing, which generally require models to be ready at launch or soon afterwards.