Detecting and classifying images of clouds at the pixel level is a crucial operation when working with satellite imagery. The presence of clouds will distort the data across all wavelengths, and u and unclassified clouds over extended periods of time will render any conclusions drawn from the data useless. The type of cloud (opaque, cirrus, etc.) is also important as different types of clouds will distort the data is different ways. Many methods have been proposed to detect and classify clouds in satellite images, but they all come short of the consistent performance needed to be used in many industry applications.

This paper will present a deep learning approach to detecting and classifying cloud types. Simple dense neural networks on single pixels have been tested by others and have shown improvements over previous methods of cloud detection. We will be exploring the use of time-series analysis when detecting the clouds, rather than only referencing a single image, to show how previous values of a given pixel can help improve model performance. Our experiments show that including time-series information can improve robustness when detecting and classifying clouds, especially on cloud shadows and cirrus clouds.