Geostationary passive retrieval of ice water path with quantile regression neural networks
| dc.contributor.author | Amell Tosas, Adrià | |
| dc.contributor.department | Chalmers tekniska högskola / Institutionen för rymd-, geo- och miljövetenskap | sv |
| dc.contributor.examiner | Eriksson, Patrick | |
| dc.contributor.supervisor | Eriksson, Patrick | |
| dc.contributor.supervisor | Pfreundschuh, Simon | |
| dc.date.accessioned | 2021-06-18T06:31:13Z | |
| dc.date.available | 2021-06-18T06:31:13Z | |
| dc.date.issued | 2021 | sv |
| dc.date.submitted | 2020 | |
| dc.description.abstract | Accurate characterisation of clouds can help determine their influence on weather hazards, climate effects, and the hydrological cycle. The retrieval of atmospheric ice mass can contribute to this characterisation and one way to quantify it is with the ice water path (IWP). The polar orbiting CloudSat satellite provides profiles of the ice in clouds, but it suffers from a long revisit period and its cross-track footprint is less than 2km. This poses a challenge in the study of atmospheric ice variability, both in time and space. Imagers aboard geostationary satellites can provide pictures of one side of the Earth in short time intervals, which can help overcome this limitation. Current algorithms to retrieve IWP from geostationary passive remote sensors are based on physical approaches or target a specific type of cloud. Quantile regression neural networks (QRNNs) can capture complex relationships for different conditional quantiles of a dependent variable. This can be used to predict a retrieval value, as well as the case-specific uncertainty. This work employs QRNNs to retrieve the IWP distribution using data from the SEVIRI instrument aboard the geostationary Meteosat-9 satellite calibrated against DARDAR, a product derived from combining CloudSat and CALIPSO measurements. The QRNNs are trained and evaluated on a large African region, covering both land and ocean areas. A multilayer perceptron and a convolutional neural network are compared, and it is seen that the use of spatial information improves both the retrieval and the associated uncertainty. Models trained using all SEVIRI channels show better overall performance in several metrics, although models that use only infrared channels show a relatively similar performance. Moreover, the retrieval with infrared channels shows to satisfactorily retrieve the IWP throughout the diurnal cycle. Models that use visible and infrared channels likely suffer an artefact in the diurnal cycle, but it cannot be completely assessed due to the coverage of the reference data. Monthly means and diurnal variations are compared with a physical-based IWP retrieval for two tropical areas, and it is found correlations that favour the QRNN models. Finally, it is suggested that the models trained on Meteosat-9 observations may also be used on observations from other Meteosat satellites, expanding the usage of the models beyond the lifetime of Meteosat-9. | sv |
| dc.identifier.coursecode | SEEX30 | sv |
| dc.identifier.uri | https://hdl.handle.net/20.500.12380/302612 | |
| dc.language.iso | eng | sv |
| dc.setspec.uppsok | LifeEarthScience | |
| dc.subject | quantile regression, neural networks, ice water path, SEVIRI | sv |
| dc.title | Geostationary passive retrieval of ice water path with quantile regression neural networks | sv |
| dc.type.degree | Examensarbete för masterexamen | sv |
| dc.type.uppsok | H | |
| local.programme | Engineering mathematics and computational science (MPENM), MSc |