Improving satellite measurements of clouds and precipitation using machine learning

Abstract: Observing and measuring clouds and precipitation is essential for climate science, meteorology, and an increasing range of societal and economic activities. This importance is due to the role of clouds and precipitation in the hydrological cycle and the weather and climate of the Earth. Furthermore, patterns of cloudiness and precipitation interact across continental scales and are highly variable in both space and time. Therefore their study and monitoring require observations with global coverage and high temporal resolution, which currently can only be provided by satellite observations. Inferring properties of clouds or precipitation from satellite observations is a non-trivial task. Due to the limited information content of the observations and the complex physics of the atmosphere, such retrievals are endowed with significant uncertainties. Traditional methods to perform these retrievals trade-off processing speed against accuracy and the ability to characterize the uncertainties in their predictions. This thesis develops and evaluates two neural-network-based methods for performing retrievals of hydrometeors, i.e., clouds and precipitation, that are capable of providing accurate predictions of the retrieval uncertainty. The practicality and benefits of the proposed methods are demonstrated using three real-world retrieval applications of cloud properties and precipitation. The demonstrated benefits of these methods over traditional retrieval methods led to the adoption of one of the algorithms for operational use at the European Organisation for the Exploitation of Meteorological Satellites. The two other algorithms are planned to be integrated into the operational processing at the Brazilian National Institute for Space Research, as well as the processing of observations from the Global Precipitation Measurement, a joint satellite mission by NASA and the Japanese Aerospace Exploration Agency. The principal advantage of the proposed methods is their simplicity and computational efficiency. A minor modification of the architecture and training of conventional neural networks is sufficient to capture the dominant source of uncertainty for remote sensing retrievals. As shown in this thesis, deep neural networks can significantly improve the accuracy of satellite retrievals of hydrometeors. With the proposed methods, the benefits of modern neural network architectures can be combined with reliable uncertainty estimates, which are required to improve the characterization of the observed hydrometeors.