Clouds in satellite imagery result in biased vegetation indices, erroneous land cover classifications, and inaccurate biomass inferences. The Climate Corporation’s deep learning based cloud mask outperforms traditional machine learning based cloud mask for multiple imagery sources and can be applied efficiently at scale.
Techniques such as traditional multi-band thresholding and machine learning have been applied to solve the various challenges of cloud detection in satellite imagery. Deep learning has revolutionized the domain of image processing. This talk will highlight our approach of using deep learning for cloud detection to accurately detect and mask various types of clouds in satellite imagery, especially for improving detection on thin/transparent clouds and reducing false detection on bright soil and urban areas. We will also illustrate the application of this model on imagery from multiple satellite sensors, as well as efficiently scaling it up to big data.