A Bayesian machine learning approach to passive microwave precipitation retrievals
| dc.contributor.author | Norrestad, Teodor | |
| dc.contributor.department | Chalmers tekniska högskola / Institutionen för rymd-, geo- och miljövetenskap | sv |
| dc.contributor.department | Chalmers University of Technology / Department of Space, Earth and Environment | en |
| dc.date.accessioned | 2019-07-05T11:53:11Z | |
| dc.date.available | 2019-07-05T11:53:11Z | |
| dc.date.issued | 2019 | |
| dc.description.abstract | A machine learning-based approach to precipitation retrievals, using Quantile Regression Neural Networks (QRNNs), is developed for data from the Global Precipitation Measurement (GPM) mission. The retrievals are conducted within a Bayesian framework where the networks are trained to predict quantiles of the posterior distribution of rain rates, conditioned on passive microwave observations. In this way, rain rates are retrieved along with the associated retrieval uncertainties. The effects of including additional spatial information as input to the QRNNs are also investigated. Different QRNNs are trained and tested, first globally over oceans and then over the U.S Great Plains. In both cases, the performance of the QRNNs are compared to the Goddard Profiling Algorithm (GPROF), a state-of-the-art passive microwave retrieval algorithm. The primary results are those over oceans, where the QRNNs show great performance on similar levels as GPROF with respect to point estimate metrics such as the mean squared error. Furthermore, the QRNN retrievals are very fast, taking less than a millisecond per footprint on a standard computer. It turns out that extra spatial information improves the QRNNs, especially on making rain-no rain classifications with fractions of true positives and true negatives exceeding 0.67 and 0.96 respectively. Furthermore, the QRNNs manage to produce well calibrated quantiles, resulting in good confidence intervals to account for retrieval uncertainties. Over the Great Plains, the results are promising but are based on much smaller amounts of data and are thus less significant. | |
| dc.identifier.uri | https://hdl.handle.net/20.500.12380/256899 | |
| dc.language.iso | eng | |
| dc.setspec.uppsok | Technology | |
| dc.subject | Geovetenskap och miljövetenskap | |
| dc.subject | Klimatforskning | |
| dc.subject | Meteorologi och atmosfärforskning | |
| dc.subject | Earth and Related Environmental Sciences | |
| dc.subject | Climate Research | |
| dc.subject | Meteorology and Atmospheric Sciences | |
| dc.title | A Bayesian machine learning approach to passive microwave precipitation retrievals | |
| dc.type.degree | Examensarbete för masterexamen | sv |
| dc.type.degree | Master Thesis | en |
| dc.type.uppsok | H | |
| local.programme | Complex adaptive systems (MPCAS), MSc |
Ladda ner
Original bundle
1 - 1 av 1
Hämtar...
- Namn:
- 256899.pdf
- Storlek:
- 3.78 MB
- Format:
- Adobe Portable Document Format
- Beskrivning:
- Fulltext