Vortrag: Detection of precipitating clouds based on optical satellite sensors using Machine Learning
Apostolos Giannakos, Alexander Jann Zentralanstalt für Meteorologie und Geodynamik (ZAMG), Vienna, Austria
This study aims at developing rainy clouds delineation schemes based on the high spectral resolution of Meteosat Second Generation - Spinning Enhanced Visible and Infrared Imager (MSG-SEVIRI). The possibility of developing precipitating cloud detection schemes was investigated, using the enhanced thermal infrared spectral resolution of the MSG satellite data. Two different rain clouds detection methods were proposed that incorporate spectral cloud parameters. The first is an Artificial Neural Network (MLP) model for rain and no rain discrimination and the second model is a Random Forest (RF) algorithm that relies on the correlation of spectral cloud parameters and rain information recorded from rain gauge stations. Both algorithms were developed using the scikit-learn python library. The rainy clouds detection schemes were trained using as rain information spatially and temporary collocated rain gauge data for several rain cases. The two rain detection schemes were validated against independent rain gauge measurements. During the training phase RF model presented the best performance among all the rain cloud delineation models. When evaluating the techniques performance against the independent rain gauge dataset RF algorithm still provides the best performance among all precipitating clouds detection schemes.
The objective of this study is to investigate the possibility of the random forests ensemble classification technique to improve rain area delineation based on the correlation of spectral and textural cloud properties extracted from Meteosat Second Generation - Spinning Enhanced Visible and Infrared Imager (MSG-SEVIRI) data with rainfall data recorded by the National Observatory of Athens (NOA) meteorological stations.
Two different precipitating cloud detection schemes are examined that use spectral cloud parameters along derived from the thermal infrared MSG satellite data to discriminate rain from no rain clouds. The first is an Artificial Neural Network (MLP) algorithm for rain cloud detection and the second scheme is a Random Forest (RF) algorithm that is based on the correlation of spectral cloud measures and rain information recorded from rain stations. The two ML approaches are implemented in python using the Scikit-learn package. The rain and no rain clouds descrimination models were calibrated using as rain information spatially and temporary matched rain gauge data for several rain events.
The rain cloud areas detection schemes were calibrated and evaluated using spatio-temporally matched 15min observation datasets of seven SEVIRI thermal infrared channels and rain gauge data. In order to create the two precipitating cloud detection models SEVIRI thermal infrared data were used and acquired a 15 min time intervals with a spatial resolution of 3x3 km2 at sub-satellite point reaching 4 x 5 km2 at the study area. Rain gauge data from 88 stations of the National Observatory of Athens were used as rainfall information to train the models. Models were validated against independent rainy days that were not used for training the rain area delineation algorithms.
- Extended Abstract
- Detection of precipitating clouds based on optical satellite sensors using Machine Learning