Abstract's details

Detecting rain cells in SARAL/AltiKa data: results from a supervised learning experiment

Pierre Prandi (CLS, France)

CoAuthors

Benjamin Pelvet (CLS, France); Julien Bocage (CLS, France); Gérald Dibarboure (CNES, France)

Event: 2022 Ocean Surface Topography Science Team Meeting

Session: Regional and Global CAL/VAL for Assembling a Climate Data Record

Presentation type: Type Poster

Operating in Ka band, SARAL/AltiKa is more sensitive to liquid water in the atmosphere which can drastically attenuate the signal even for moderate events (eg light rain) and impact the retrieval of geophysical parameters. This turned out not to pose any problem regarding data availability thanks to the instrument's link budget margins. The operational products come with a rain flag based on the pre-launch work of Tournardre et al. (2009), which was finely tuned during the cal/val phase of the mission (Tournadre et al., 2015).

This rain flagging algorithm is based on the detection of along-track short scale coherent variations of the off nadir angle derived from the trailing edge slope of the waveforms. This algorithm is efficient at detecting rain events, sigma blooms but is also sensitive to other perturbations of the waveform (true mispointing events for example).

In this analysis we try to train machine learning/artifical intelligence classifiers to detect rain events in SARAL/AltiKa data. We tested a range of supersived learning methods on a database of colocations between SARAL/AltiKa and rain rates derived from SSMI-S F16, F17 and WindSat.
This was done first on 1Hz data with a range of algorithms (K-Nearest Neighbors, Random Forests) and input features (altimeter parameters, radiometer parameters).
On 40Hz data, more complex architectures (Convolutionnal Neural Nets) were tested on waveforms alone and on waveforms plus other parameters (eg brightness temperatures).

Results show that both at 1 and 40 Hz we are able to train efficient detectors of rain events, even when waveforms are the only input available. Removing measurements that are detected as 'rainy' provides similar mission performance improvements compared to the current rain flag (measured at cross overs or through SLA PSD estimation) while removing less measurements.

While this work remains preliminary, this class of methods could be useful the editing of current and future altimetry missions.
 

Poster show times:

Room Start Date End Date
Mezzanine Tue, Nov 01 2022,17:15 Tue, Nov 01 2022,18:15
Mezzanine Thu, Nov 03 2022,14:00 Thu, Nov 03 2022,15:45
Pierre Prandi
CLS
France
pprandi@cls.fr