Bad weather in Austria again? You should make a short trip to catch the sun. But be aware – more than 55% of the Earth’s surface is always clouded!
Machine Learning is a powerful approach for Big Data problems where human intelligence alone isn’t sufficient.
Do you have a Machine Learning problem? Or don’t you know whether your problem can be solved with Machine Learning? Don’t struggle with it. Contact us. We look forward to cracking the nut!
We have presented CI4Clouds at the EUMETSAT Meteorological Satellite Conference in September 2016 in Darmstadt.
What is CI4Clouds?
The Cloud Problem
Satellites continuously provide us with optical data in several wavelengths of the Earth’s surface. The problem is to detect clouds. A misclassification has a strong impact on further algorithms of a processing chain, which assume clear sky conditions. 55% of the Earth surface is always clouded.
The project Computational Intelligence 4 Clouds – CI4Clouds – uses satellite pictures, like from the Korean GOCI satellite instrument, to identify clouds. GOCI, the Geostationary Ocean Color Imager, was built for the purpose of ocean surveillance around Korea, Japan, and Eastern China. More than 2 billion people live in this area. Catalysts’ project partner is the Austrian meteorological office – ZAMG.
CI4Clouds – The Catalysts Approach
We set up a number of machine learning algorithms which compete against each other: Deep Learning (DL) and Convolutional Neuronal Nets (CNN), Random Forests (RF) as well as Support Vector Machines (SVM).
For training high performance hardware is used. The evaluation is against existing cloud masks from GOCI but also European and US satellite cloud products.
RF is the winning algorithm. RF also performs better than the known cloud products. Domain-specific pre- and postprocessing enhances results even further.
This project is carried out in a partnership of ZAMG and Catalysts. CI4Clouds is funded by the Austrian Federal Ministry of Transport, Innovation and Technology (BMVIT) under the program “ICT of the Future” between 2015 and 2016. More information