In recent years machine learning and especially deep learning has gained enormous momentum through its fundamental breakthroughs in image recognition and language processing. Even though the underlying data properties of those applications differ from tropospheric ozone time-series or satellite data, the methods and concepts of such systems can help to distil information from existing data. This information will allow the TOAR-II community and others to enhance their understanding of tropospheric ozone. Within this working group (WG), we will focus on knowledge that can be derived from the TOAR database, in conjunction with other data, to shed light on existing scientific problems and questions.
Ozone formation and loss processes are highly nonlinear and heavily depend on other chemical variables and meteorological properties. Air pollutant concentrations are controlled by four main types of processes: emissions, transport, (chemical and physical) transformations, and loss processes such as deposition and washout. These processes that affect production and removal of ozone are of a complexity where high-quality interpolation, forecasting, and quality assurance would benefit from ML approaches. This WG will bring together ozone and machine learning experts, allowing us to develop best practice guidelines on data preprocessing and model evaluation with respect to AI applications.
Key objectives of the WG include:
-
Contribution to the assessment of the Phase II of the TOAR database
-
Documenting and developing best practices for handling TOAR data for use with machine learning applications, and contribute to the development of a benchmark dataset
- Identify optimal periods for training, validation and testing data
- Possible uses: an unseen data set for ML model assessment
-
Provide best efforts to develop methods for interpolation of missing values in the TOAR database. Suggested methods:
- derive reliable information on air quality at stations where incomplete or no measurements exist (spatial interpolation to "new" or "unknown" station locations;
- fill temporal gaps in the measurement time series at individual station locations by identifying a transfer function
-
Co-ordinate and lead a paper for the TOAR-2 Special Issue. Describe/assess and provide benchmakrs where appropriate for
- (surface) ozone forecasting
- preparation of interpolated/fused dataset that attempt to provide better consistency / accuracy / coverage of atmospheric composition
- use of ML methods in atmospheric chemical modelling
-
Liaise with other TOAR-2 WG e.g. statistics and health on issues of mutual interest