A machine learning based 24-h-technique for an area-wide rainfall retrieval using MSG SEVIRI data over Central Europe
The aim of the present study was to develop a 24-h-technique for the process-related and quantitative estimation of precipitation in connection with extra-tropical cyclones in the mid-latitudes based on MSG SEVIRI data using the machine learning algorithm random forest. The algorithms and approach...
|Online Access:||PDF Full Text|
No Tags, Be the first to tag this record!
|Summary:||The aim of the present study was to develop a 24-h-technique for the process-related and quantitative estimation of precipitation in connection with extra-tropical cyclones in the mid-latitudes based on MSG SEVIRI data using the machine learning algorithm random forest.
The algorithms and approaches needed were successfully developed and implemented within three working packages:
(WP1) The cloud property retrieval SLALOM, first developed for Terra MODIS, was successfully transferred and adapted to the specific requirements of the SEVIRI system and an extensive validation study was carried out. The cloud optical properties retrieved by SLALOM, namely cloud effective radius and cloud optical thickness that were needed for satellitebased rainfall estimation in WP2 and WP3, were compared against the well known and validated NASA MODIS cloud property product (MODIS 06) as well as the cloud optical depth product (2B-TAU) of CloudSat. The suitability of SLALOM has been shown over the North Atlantic and over the European continent (chapter 3).
(WP2) A new 24-h-technique for rainfall rate assignment was developed for MSG SEVIRI using the machine learning algorithm random forest as fundamental prediction algorithm. Based on the precipitation processes in connection with extra-tropical cyclones, rainfall rates were assigned to advectivestratiform and convective precipitating areas by means of individual RF models. As predictor variables for the RF models satellite-based information on cloud top height, cloud top temperature, cloud phase and cloud water path were chosen. The different illumination conditions (daytime, twilight and night-time) were taken into account with a proper SEVIRI spectral channel selection as surrogates for theses cloud physical parameters. The development was realised in three steps: First, an extensive tuning study was carried out to customise each of the RF models. Secondly, the RF models were trained using the optimum model parameter values found in the tuning study. Finally, the final RF models were used to predict rainfall rates using an independent validation data set and the results were validated against co-located rainfall rates observed by the RADOLAN RW product of the DWD. The outstanding validation results during all times of the day confirmed the ability of RF as tool for the rainfall rate assignment technique from MSG SEVIRI data (chapter 4).
(WP3) A new coherent daytime, twilight and night-time rainfall retrieval was developed for MSG SEVIRI. The technique aims to retrieve rainfall rates for precipitation events in connection with extra-tropical cyclones in the midlatitudes in a continuous manner resulting in a 24 hour prediction. Based on the dominant precipitation processes, the proposed rainfall retrieval consists of three steps which are applied consecutively by means of individual RF models to get the final product: (i) Identification of precipitating cloud areas. (ii) Separation of precipitating areas into predominately convective and advective-stratiform cloud regions. (iii) Individual process-oriented assignment of rainfall rates to these cloud areas. Again, the relationship between cloud top temperature, cloud top height, cloud water path and cloud phase was used to retrieve information about precipitation and according to the illumination conditions, a suitable selection of the predictor variables were taken into account as input to the RF models (chapter 5).
The newly developed rainfall retrieval technique was tested in an extensive validation study over Germany using the radar-based RADOLAN RW product as reference data. The validation results show reliable performance of the new technique concerning rain area detection, rain process separation as well as rainfall rate assignment during all times of the day which enables the estimation of precipitation for 24 hours of a day. Hereby, the twilight applicability of the technique as well as good rainfall rate prediction performances even on an hourly basis are particularly remarkable and set this study apart from other rainfall retrievals. For the first time, a 24-h precipitation monitoring becomes possible for precipitating clouds of not only convective but also of advective-stratiform character, opening many areas of application.|