Case study of machine learning using meteorological data

Keywords: MSG, SEVIRI, Meteorological Radar, Precipitation, Earth Observation, Machine Learning

TL;DR: Case study of Empirical Algorithms and Machine Learning methods used to estimate rain rate in specific moment using data from MSG satellite.

Abstract: Paper shows case study of machine learning carried out on meteorological data. The main goal in presented case study is to estimate rain rate over certain area using satellite data from Meteosat Second Generation (MSG). As a reference for ML process the data from network of meteorological radars located in Poland is used. Input and reference data had to be geometrically corrected and collocated before feeding it to ML process. In some variants training data was prepossessed using aggregation, which purpose was to perform data generalization. The subjects of the ML process were two empirical algorithms: Vicente and Roebeling. In presented case study also were used ML methods such as: shallow neural networks in two variants, decision trees and random forest. Experiments were conducted using data from 2015 as a training input, and data from 2016 for evaluation process. Empirical algorithms and ML models predicting rain rate in mm/h did not give satisfactory result. Training ML models to predict rain rate as order of magnitude -- dBZ, gave better results than predictions in mm/h. However there is still room for improvements. Training empirical algorithms and ML methods for both mm/h and dBZ prediction on aggregated data performed much better than for unaggregated data, but tests performed after training resulted with slightly worse statistical metrics values.

Serve As Reviewer: ~Børre_Bang1

Submission Number: 36