Photo by Corina Behrens
Publish Date: 17.01.2022
Category: Interdisciplinary research, Our contribution to sustainable development goals
Sustainable development goals: 9 Industry, innovation and infrastructure, 11 Sustainable cities and communities, 13 Climate action, 14 Life below water, 15 Life on land, 17 Partnerships for the goals (Indicators)
An interdisciplinary group of researchers from the University of Ljubljana, Faculty of Computer and Information Science (Mag. Lojze Žust and Prof. Matej Kristan), the National Institute of Biology (Dr Matjaž Ličer) and the Slovenian Environment Agency (Anja Fettich) has developed a deep neural network for predicting flooding in the northern Adriatic that can compete with physical numerical models.
Figure 1: HIDRA logo
Due to global warming, the flooding of coastal towns is an increasingly pressing problem. According to the model projections, the Slovenia’s median sea level will rise by 30 – 100 cm by the year 2100. As a result, coastal towns such as Piran will be subject to flooding as much as twice a day. Precise prediction of flood events is therefore vital for timely warning and protection of the coastal population and economy.
Figure 2: With a 100 cm rise in the sea level, the area of the Slovenian coast marked in red will be entirely flooded. Source:
However, the prediction of coastal flooding in the northern Adriatic is extremely challenging due to the interaction of atmospheric phenomena, the specific topographic features of air flow and sea currents and the resonance properties of the Adriatic basin. Prediction models thus rely on simulation of coupled atmospheric and oceanic models, which is slow and subject to modelling errors.
A group of researchers from UL FRI, NIB and ARSO tackled the challenge of flood prediction using machine learning. They developed a deep neural network HIDRA, which predicts the sea level at a selected point for the next 72 hours from a sequence of atmospheric images and past sea level measurements. The network outperforms the most advanced operational ocean model NEMO in accuracy, while being half a million times faster. Running on an ordinary laptop, HIDRA does not require specialized hardware, and generates half a million times smaller energy footprint. It can therefore be used by ordinary people and local communities. This is the first machine-learning method that outperforms oceanic physical numerical models on this task. HIDRA is already a part of the operational predictions provided by the Slovenian Environment Agency, and its predictions can be accessed at: https://lojzezust.github.io/hidra-visualization/sl/. HIDRA was presented in an article published in a prestigious journal of geophysics Geoscientific Model Development, and it was selected as one of the most outstanding research achievements in the field of Natural Science and Technology, with the designation “Excellent in Science 2021”. The selection is made every year by the Slovenian Research Agency.
Figure 3: HIDRA predicts the sea level at a selected point for three days into the future with an accuracy comparable to that of numerical models. Source