A detailed look at the state of Flood Forecasting in climate science and the integration of Machine Learning Pipelines
Floods are some of the most common and devastating natural disasters in the world. As a Canadian citizen, it was concerning to me finding out that floods are the most common natural hazard in the country and among the costliest. A recent report showed that in the last decade, flooding has killed on average more than 5000 people a year around the world. For this reason, it is extremely important that we develop innovative technologies to mitigate the effects of this devastating natural disaster. Early warnings can prevent up to 43% of fatalities, and 35% of economic damages due to flooding, as anticipated floods allow for precautions to be taken and people to be warned so that they can be prepared in advance for flooding conditions, avoiding potential fatalities.
This article will review the state of flood forecasting and how machine learning can be integrated to improve the results of flood models.
Why do floods occur?
To be able to accurately forecast floods, we first need to understand the science behind flooding.
There exist 3 types of floods, each with their own causes:
- Surface (Pluvial flood)
- River (Fluvial flood)
- Coastal (Surge flood)
Surface (Pluvial Flood)
A pluvial, or surface water flood, occurs when heavy rainfall creates a flood independent of an overflowing water body. A common misconception about flooding is that you must be located near a body of water to be at risk. However, pluvial flooding can happen in any location — urban or rural — even in areas with no water bodies in the vicinity. Pluvial floods occur gradually, which provides people with time to go indoors or leave the area. The level of water is low to the ground (rarely more than one meter) and causes no immediate threat to lives. However, depending on the flooded area, it may cause significant economic damage.
River (Fluvial Flood)
Sometimes also called riverine floods, fluvial floods occur when a river overflows its bank. They are characterized by a slow and steady rise of the water level, which leads to an eventual flood. The water level rise could be due to excessive rain or snowmelt.
A coastal flood, as the name suggests, occurs in areas that lie on the coast of a sea, ocean, or other large bodies of open water. Common causes of coastal flooding are high tide, tsunamis and storm surge. Storm surge — produced when high winds from hurricanes and other storms push water onshore — is the leading cause of coastal flooding and often the greatest threat associated with a tropical storm.
Flash Floods (4th type)
There isn’t a global definition of what a flash flood is, as some experts categorize it as a type of river flood, while others categorize it as a type of pluvial flood. Hence, the definition of flash floods is very context and place-specific. However, I still wanted to put the term in here as you will be hearing the term a lot if you continue research in flooding. What you need to know is that flash floods are the most devastating in nature as they can arrive in a “flash”.
The common agreement is that it is characterized by heavy rain falling in the area during a very short time. They are the most common flood type in normally-dry channels in arid zones, and occur within minutes to hours after a heavy rain event and produce raging torrents of water that move with great speed.
Flash floods are characterized by an extremely fast rise in water level, leaving little to no time for the population to react appropriately. These are the deadliest type of flood because they are often unexpected and underestimated by the population. People often forget the growing heavy debris that is generated by the flash flood, which can collide into buildings and objects with immense force.
How are floods predicted? (models and real-life analysis)
Now that we understand the different types of flooding and how each of them is caused by different factors, it is easy to see how a universal flood model cannot be created to forecast potential floods. For each type of flood, we need different models using different predictors. I will be explaining the general methodologies that go with the prediction of each type of flood.
In general, when we forecast floods, what we really want to forecast is how the flood is going to impact the human population. We want to be able to pinpoint exactly which areas are going to be affected by the flood, as well as determine how damaging the flood will be on the infrastructure and surrounding landscape, and calculate the potential economic loss associated with this flood.
Predicting surface water/Pluvial floods
Since surface water floods are exclusively caused by heavy rain, we can predict pluvial floods by forecasting the quantity of precipitation. Other factors, such as the rate at which the drainage system can intake water and the type of soil matters too; for example, water doesn’t move very well through clay soils, so less rainfall will be absorbed into the ground and more of it will runoff above ground.
However, while there are models to help predict the behaviour of rainstorms, they are not always accurate. Predicting where rainfall is the heaviest is not as straightforward as you may think; for example, thunderstorms are complex systems that are unpredictable and can change pretty quickly. Trying to understand how heavy rainfall will impact people involves understanding the particularities on the ground.
Predicting Coastal Floods
As the population continues to grow and the number of coastal cities rises every year, coastal floods become increasingly a risk for these individuals. Being able to accurately forecast these types of floods becomes crucial to mitigate and avoid the economic damages and loss of human lives. Coastal floods are predicted by analyzing the behaviour of the sea.
Predicting Riverine Floods
This is my area of expertise and research, so I will be going the most in detail with these kinds of floods. Remember that riverine floods are caused by an overflow of a river’s banks. Hence, to predict riverine floods, we can divide the task into three steps:
- Extracting the data necessary for the models
- Predicting the discharge of the river (Hydrologic model)
- Predicting which areas the water will flood (Hydraulic model)
1. Data extraction
This first step in building a flood model can easily be overlooked. However, it is often the most difficult part of the task, as data can be extremely difficult to obtain.
For example, since the flow of the river (scientifically called the “discharge” and measured in [m^3/s]) is one the few accurate indicators of potential flooding (fluvial flooding events always show a dramatic increase in the river’s discharge), we want to be able to extract that data if possible. It is so accurate that when the discharge exceeds a certain quantity, we can almost be certain that there will be flooding in the nearby area. We can visualize this through a hydrograph.
As you can see in the hydrograph above, the increase in discharge is often correlated with an increase in precipitation, with a lag time in between. Precipitation is thus a crucial factor and important data to be extracted in order to create an accurate flood forecasting model.
However, the problem arises when in many regions around the world, that form of data is unavailable, as the discharge can only be measured through instruments called steam gauges, which are hard to build and maintain. In fact, currently, there are more stream gauges being dismantled than being built every year. As a result, scientists often need to use a scientific model called the hydrologic model to calculate the river discharge rather than directly measuring it.
2. Hydrologic model
As stated previously, the purpose of the hydrologic model is to predict the discharge of the river based on different geological, environmental and meteorological processes in the area. The hydrologic intakes various parameters as you can see in the image above. All of the modelling of the river discharge value is based on scientific processes that have been researched and proven to be extremely accurate based on theory.
3. Hydraulic model
The hydraulic model is what will allow us to create a detailed map of which areas are going to be flooded. To do this, we need to provide the model with 2 parameters, the discharge of the river as well as what we call a surface elevation map, which as the name implies, provides information on the elevation of the different areas. This produces the end result that we desire, as we are able to see exactly which areas are going to flood and allowing us to properly warn the population.
How accurate are flood forecasts?
The problem that arises as scientists attempt to build this map is that the quality of the data becomes often very distorted. This is also extremely evident if we use google maps. Zooming in to a big city like New York allows us to see different buildings on the resolution of 1x1 m, while when we look at other parts of the world, the quality of the image is extremely low.
Another issue that arises as a result of the bad data can be seen in the surface area that models produce. Using traditional satellite imagery, the topography is extremely expensive and difficult because we need to send a specialized satellite into space, which becomes extremely expensive, sending waves and calculating the elevation of the ground by measuring the time it takes for the wave to come back. Not only are these measurements lacking resolution, but one must realize that because of how expensive the mission is, scientists are not able to regularly update the surface elevation map every year. Moreover, the Earth and its river change drastically every year. Here’s a video below in case you didn’t know that rivers can move.
Flood forecasts are just not that precise at the moment. While we can generally have a good idea of exactly when flooding will be happening, we cannot accurately pinpoint the outline of the flood. Consequently, warnings for the population are often futile, as they have no idea as to where to escape and hide in their homes, underestimating the hazardous flood that might take away their flood.
How can we improve flood models?
The classical approach to flood forecasting, with hydrologic models, has always been to use a list of physical processes that, given enough data, can yield good results.
We can use AI to produce a surface elevation map. We have tons of satellites in the sky which are constantly taking pictures of the Earth. Combining these images taken by different satellites at different positions of the same area on the Earth can allow us to produce a surprisingly detailed map of the elevation since we know exactly the angle at which the satellite took the image.
Another way ML can intervene is in building the flood models themselves. While we have gotten to the point where science has advanced so far that we can exactly model the physical processes in the world around us with mathematical equations, we simply lack in many cases the data to calculate this. Machine learning can intervene and replace these models that require tons of data, and yield similar if not better results than traditional models.
The future of Flood forecasting
Once we are able to build accurate enough models with little to no error, we need to effectively bring this to human society by notifying affected populations in a timely and effective manner: this is what we call the delivery.
The most notable case that I have seen based on the research I have done is the project led by a team at Google, with the goal to bring flood warnings in an effective manner to the population at risk. They have implemented a warning system with colour codes mapping out the risks in different areas.
However, still a lot of work needs to be done, considering thousands still continue to die every year because of flooding. Keeping this in mind, I’ve been working on a flood forecasting project of my own in order to properly warn populations. My goal is to make this flooding information accessible to the public, as the only resources available on the internet at the moment is the USA NOAA website, which produces 14-day forecasts, as well as the GloFAS (Global flood awareness system) which produces a detailed map of the world and impact calculation.
Thank you for reading,