Using Machine Learning to Detect Cloud Signatures in COSMIC-2 Radio Occultations
Abstract
The goal of this study is to infer cloud presence using retrievals of the neutral atmosphere by radio occultation (RO) from the constellation of Global Navigation Satellite Systems. RO is well known for providing very high vertical resolution measurements of microwave refractivity. However, cloud water fields themselves contribute weakly to microwave refractivity, but they do leave other signatures in refractivity profiles that are commonly linked to their presence. These include refractivity enhancements brought on by large amounts of water vapor and/or altered temperature lapse rates. When accompanied by forecasts of temperature and density from numerical weather prediction, RO retrievals of refractivity and "dry temperature" are expected to contain information on the presence of clouds that can be mined using modern machine learning (ML) approaches. We are assembling a large collection of RO profiles along with coincident NWP forecast fields of temperature, water vapor and cloud water, and corresponding contemporaneous GOES-16 Advanced Baseline Imager (ABI) sensor data sets. Cloud truth fields are established for each RO profile using a hybrid cloud modeling technique that exploits the information content of the ABI and GFS cloud-water datasets. Coupled with refractivity and dry-temperature observations from the second-generation Constellation Observing System for Meteorology, Ionosphere and Climate (COSMIC-2), we are designing ML techniques that categorize the cloud state within RO profiles. The ML model is being designed to match ABI-based cloud observations (i.e., cloud/no-cloud, and perhaps other properties including cloud fraction) from corresponding RO profiles. We are experimenting with multiple ML approaches including the xgboost algorithm and custom-built deep neural networks in order to probe the potential of the RO data to diagnose cloud properties. Initial decision-tree experiments show strong promise, explaining up to 70% of the variability in GFS forecast relative humidity based on RO observables. This is a first step toward cloud detection, since high RH is a reasonably strong indicator of cloud presence.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2020
- Bibcode:
- 2020AGUFMA059.0016C
- Keywords:
-
- 3311 Clouds and aerosols;
- ATMOSPHERIC PROCESSES;
- 1906 Computational models;
- algorithms;
- INFORMATICS;
- 1914 Data mining;
- INFORMATICS;
- 1922 Forecasting;
- INFORMATICS