Addressing Cloud-Height Retrieval Improvements with Convolutional Neural Networks

Proper cloud height retrieval still faces difficulties in complex settings, such as multi-layer cloud decks and thin cirrus imagery. This has an important negative effect on atmospheric motion vector (AMV) wind retrievals among other issues. However, current retrieval techniques are mostly based on individual pixels, and the use of the surrounding context of the retrieval is limited. Here, convolutional neural networks can play an important supporting role in relating the quasi-subjective properties of the image data nearby in space and time. CALIPSO/CALIOP retrievals are used to train the model in a large multi-year dataset. We explore the impact of this additional information on cloud height retrievals as well as its effect on subsequent AMV retrievals.