Predicting the level of water in rivers or streams can prove to be invaluable in areas with a high risk of flooding or droughts. While traditional models are based primarily on hypotheses about processes, another approach is emerging: artificial neural networks. Anne Johannet, an environmental engineering researcher at IMT Mines Alès, uses this approach.
This article is part of our dossier “Far from fantasy: the AI technologies which really affect us.”
In hydrology, there are two especially important notions: high and low water levels. The first describes a period in which the flow of a watercourse is especially high, while the second refers to a significantly low flow. These variations in water level can have serious consequences. A high water level can, for example, lead to flooding (although this is not systematic), while a low water level can lead to restrictions on water abstraction, in particular for agriculture, and can harm aquatic ecosystems.
Based on past experience, it is possible to anticipate which watercourses tend to rise to a certain level in the event of heavy precipitation. This approach may obtain satisfactory results but clearly lacks precision. This is why Flood Forecasting Services (SPC) also rely on one of two types of models. The first is called “reservoir modeling”: it treats a drainage basin like a reservoir, which overflows when the water content exceeds its filling capacity. But forecasts made based on this type of model may contain major errors, since they do not usually take into account soil heterogeneity or variability of drainage basin use.
The other approach is based on a physical model. The behavior of a studied watercourse is simulated using differential equations and field measurements. This type of model is therefore meant to take all the data into account in order to provide reliable predictions. However, it reaches its limits when faced with high variability, as is often the case: how land reacts to precipitation may depend on human activity, type of agriculture, seasons, existing vegetation etc. As a result, “it is very difficult to determine the initial state of a watercourse,” says Anne Johannet, an environmental engineering researcher at IMT Mines Alès. “It is the major unknown variable in hydrology, along with the unpredictability of rainfall.” Therefore, the reality may ultimately conflict with forecasts, as was the case with the exceptional rising of the Seine in 2016. Moreover, certain drainage basins are little addressed by physical models due to their complexity. The Cévannes is one such example.
Neural networks learn independently
Anne Johannet’s research focuses on another approach which offers a new method for forecasting water flow: artificial intelligence. “The benefit of neural networks is that they can learn a function from examples, even if we don’t know this function”, explains the researcher.
Neural networks learn in a similar way as children do. They start out with little information and study a set of initial data, by calculating the output in a random way and inevitably making mistakes. Then, numerical analysis methods make it possible to gradually improve the model in order to reduce these errors. In concrete terms, in hydrology, the objective of neural networks is to forecast the flow a watercourse or its water level based on rainfall. A dataset describing all the past observations about the basin is therefore used to train the model. While it is learning, the neural network calculates a flow based on precipitation, and this result is compared to real measurements. This process is then repeated several times, to correct its mistakes.
A pitfall to be avoided with this approach is that of “overlearning.” If a neural network is “overtrained,” it can eventually lose its extrapolation quality and settle for knowing something “by heart.” To give an example, if the neural network integrates the appearance of a major rise in water level on 15 November 2002, overlearning can lead it deduce that such an event will occur every year on 15 November. To avoid this phenomenon, the dataset used to train the network is divided into two subsets: one for learning and one for validation. And as the errors are corrected on the training dataset its ability to generalize is verified using the test dataset.
The main benefit of this neural network approach is that it requires much less input data. A physical model requires a large amount of data, about the nature of the land, vegetation, slope etc. A neural network, on the other hand, “only needs the rainfall and flow at a location we’re interested in, which facilitates its implementation,” says Anne Johannet. This leads to lower costs and provides quicker results. However, the success of such an approach relies heavily on rainfall predictions, which are used as input variables. And this precipitation remains difficult to forecast.
Clear advantages but a controversial approach
Today, Anne Johannet’s models are already used by public services including the Artois-Picardie Flood Prevention Service (in the Hauts-de-France region). Based on rainfall prediction, agents establish scenarios and study the consequences using neural networks. Depending on the type of basin — which may react more or less quickly — they are also able to make forecasts several hours or even a day in advance for high water levels, and several weeks in advance for low water levels.
This data can therefore have a direct effect on local authorities and citizens. For example, predicting a significant low water period could lead water supply management to switch to an alternative water source, or could lead the authorities to prohibit water abstraction. Predicting a high water period, on the other hand, could help anticipate potential flooding, based on the land structure.
Longer-term projections can also be established using data from the IPCC (Intergovernmental Panel on Climate Change). Trial forecasts of the flow of the Albarine river in the Ain department have been carried out, up to 2070, to assess the impact of global warming. The results indicate severe low-water levels in the future, which could affect land-use planning and farming activity.
However, despite these results, the artificial intelligence approach for predicting high and low water levels has been met with mistrust by many hydrologists, especially in France. They argue that these systems are incapable of generalizing due to climate change, and largely prefer physical or reservoir models. The IMT Mines Alès researcher rejects these accusations, underscoring the rigorous validation of neural networks. She suggests that the results from the different methods should be viewed alongside one another, evoking the words of statistician George Box: “All models are wrong, but some are useful.”
Article written for I’MTech by Bastien Contreras