This website uses cookies to ensure a better user experience.

To get more information, please read our Cookie Statement.


Contrast Dysl
Special — 06/07/2021

Awarded solution to the EVA2021 Data Challenge

Dr. sc. Tomislav Ivek (Institut za fiziku) and doc. dr. sc. Domagoj Vlah (Faculty of Electrical Engineering and Computing, University of Zagreb) have won the first prize at the Data Challenge of the Extreme Value Analysis 2021 conference organized by The University of Edinburgh, UK on 28.06-2.07.2021. This year the goal of the competition was to provide an accurate reconstruction of the count and burnt area of wildfires in the USA based on incomplete observation data during a period of 23 years.

Wildfires are uncontrolled fires of combustible materials mostly composed of natural vegetation. They represent an extraordinary hazard to human lives, environment, and property. Unfortunately, it is expected that wildfires will become drastically more frequent and dangerous as the global climate changes. That is why this year’s EVA competition focuses on predicting of the distribution of wildfire counts and the damaged area in locations where this data is unobserved or otherwise absent, based only on the wildfire and terrain information on other geographical locations.

The approach by T. Ivek i and D. Vlah was awarded first place with the best accuracy in predicting both burnt area and counts of wildfires. Their methodology is based on an advanced kind of so-called Variational Autoencoders, neural network models which can be trained to reconstruct noisy or damaged information. As opposed to competing methods, no expert input or data modeling is required since the neural net learns the underlying dependence between data points on its own. This particular technique can also be applied to other areas of scientific research which rely on spatio-temporal features. The authors plan to present their model in a future paper which is currently in preparation.


IF Ⓒ 2017