Pattern Recognition in Satellite Data
May 2013 – January 2014
To investigate the characteristics of tropical rainfall near coasts developed the first of its kind OpenCV based pattern recognition technique. The algorithm performs Canny-Edge detection in satellite data to separate raining contours. It then consideres different image segment moments to identify rainfall patterns that are associated with coastal land-see interaction.
The algorithm and the results of the its application to CMORPH satellite data have been published in an article in Journal of Climate.
The average diurnal rainfall cycle (1998 - 2016) shows characteristic spatio-temporal variations in coastal areas (here Indonesia) that are related to land-sea interaction.
To filter these characteristic patterns an objective recognition algorithm is developed and applied to three hourly satellite scans (snapshot 2nd Feb. 2016).
With help of the open-source lib. OpenCV the algorithm considers contour moments and contour positions to filter characteristic rainfall patterns that are related to coastal land-sea interactions (filter result).
Tracking of Servere Weather Events
March 2018 – present
My work in the Centre of Excellence of Climate Extremes focuses on the investigation of extreme rainfall events. To study the life cycles of high impact weather I am part of a team of researchers who are developing an algorithm that is able to track rainfall data of various kinds.
The algorithm considers phase correlations to predict cell movement and is able to handle splitting and merging of different storm systems. In the team it is my responsibility to add various statistics and make the algorithm work with various datasets like satellite rainfall retrievals and high resolution model output and radar data.
Application of the storm-system tracking for data captured by a research ground-based radar in Darwin, Australia
Traces of the tracked storm-systems. The tracks are used to study the life-cycle of organised convective extreme events.
Intentionally only developed for radar application I adopted the algorithm for application to any kind of binary data. Like shown here for satellite based rainfall retrievals.
Machine Learning for Rainfall Extremes
March 2018 – present
To gain a deeper understanding and also to improve the predictive skill of rainfall extremes in the Tropics is project aims at applying state of the art Machine Learning algorithms like support vector classifier or convolutional neuronal networks to investigate and predict the occurrence of extreme rainfall events based on large-scale precursor.
The different algorithmic approaches (supervised and unsupervised) will be the corner-stone for a new generation of modelling approaches of tropical convection in global numerical weather prediction and climate models.
Due to highly non-linear processes extreme rainfall events are notoriously hard to predict from large-scale precursors that are represented in forcasting models
In this ongoing project the ability of various Machine Learning approaches to predict extreme events in the tropics is tested (here test with random forest classifier)
Results from unsupervised machine learning methods (like here a dendrogram for a cluster analysis) can give insights for to better understand extreme rainfall events
A Stochastic Multicloud Model
October 2015 - August 2016
In the project I developed a coastal version of a Stochastic Multi Cloud Multi Could model. This stochastic spin-flip model calculates transition probabilities with help of a conditional Markov Chain. The results of this project have been published in Journal of Advances in Modelling Earth Systems.
The new modelling approach can help to improve the representation of tropical convection, a major source of climate projection uncertainties, in global numerical weather prediction and climate models.
The Stochastic Model (SMCM) calculates transition probabilities between three different cloud types and clear-sky with help of a conditional Markov-Chain
The coastal version that I developed is able to represent the spatio-temporal variability of clouds along tropical clouds (here shown by an idealized test-run; white are cloudy areas black clear sky)
When applied to real-world conditions over tropical coastal locations the new model show an improvement of the prediction of clouds and convection compared to traditional modelling approaches.
A Sea-Breeze Algorithm for Global Models
March 2017 - February 2018
During my final year at Monash University I designed and implemented a decision algorithm that determines the presence of local sea-breeze circulation into the latests version of the UK Met-Office's global numerical weather forecasting model.
While the ideas encapsulated in this algorithm are straightforward to implement the challenging part was the implementation of the parallel computing paradigm numerical models make use of in the typical in super computing environments. This first of its kind sea-breeze model is currently applied by the UK-Met Office and NASA's Global Modelling and Assimilation Office.
Storm systems along the Kenyan coastline caused by sea-breeze circulations. As global (coars resolution) models do not capture theses circulations, they are error prone in coastal areas.
The developed algorithm filters the models coastal area from the static land-sea mask and the dynamic sea-ice cover (a) and also applies different filters to various atmospheric fields (b).
The calculated field of not resolved sea-breeze circulations is soley based on resolved conditions and informs other parts of the climate model to help improving the prediction of rainfall in coastal areas.
Optimizing Weather Forecast Output with Observations
October 2014 - present
The aim of the project is to improve the output of numerical weather prediction (NWP) by applying an optimisation algorithm (3D variational analysis). The variational analysis nudges NWP model output to atmospheric observations from satellites, radio sondes and radars while constraining budgets conservation of various atmospheric variables.
The output of this optimisatoin process can be used to initialise high resolution cloud resolving models and large-eddy simulations.
The algorithm tries to improve foretasted variables (like temperature) nudgeing them towards observations under the constraints that some important atmospheric budgets are conserved.