Millions of people are affected by natural disasters each year. Wildfires, floods, tornadoes, volcano eruptions, we can continue this list for a long time. Some disasters can last a few seconds, another can last for weeks. But their effects can be felt for decades or even longer, and impact the global economy, infrastructure, agriculture, and human health. The worst part is that future impact of disasters will grow dramatically due to climate change. Some regions that previously rarely suffered floods or wildfires now regularly overcome the effects of these natural disasters.
Researchers collected a huge amount of data and developed models that predict disasters, but most of these models are far from perfect. Also, the problem is the amount of data that is monitored by satellites, various ground sensors all over the world each minute. It is abnormally huge with a lot of parameters. It's good that we have more information, but data needs computational resources. If we have more data, it means that our computational models became more complex and slower. Since, sometimes only a few minutes’ notice of flood or wildfire can protect people, the predictive models should work and do corrections in real time.
Artificial Intelligence (AI) techniques and approaches, like data mining, machine learning, and deep learning, can assist in disasters predictions. It is possible for AI to find hidden dependencies in data, that can be a basis for better understanding the mechanism of disasters, and, as a result, better predictions. Good predictions and warnings reduce economic losses and save lives. We can’t stop most of the disasters, like floods, hurricanes, volcano eruptions, but we can be prepared for them.
In this article, we’ll consider how data science can help in predicting different natural disasters.
Wildfires
Most countries all over the world are faced with uncontrolled fires in rural areas. We can even say that something is always burning on Earth. Of course, there are some regions that are more sensitive to wildfires. Their climate dynamics contains fire seasons. If this is not taken into account, the consequences can be horrible. Wildfires are very destructive in many ways.
- Thousands of square kilometers of forests are burned each year only in the USA. 17,939 square kilometers were burned in the USA during January-September 2019.
- Wildfires generate large amounts of smoke pollution. Fire emissions include carbon dioxide, carbon monoxide, organic chemicals, nitrogen oxides, atmospheric mercury. Smoke can travel hundreds of kilometers and impact the health of millions of people. Another danger is that a big fraction of smoke is the greenhouse gases.
- Wildfires affect infrastructure, agriculture, and humans’ lives. They are causing losses of billions of dollars every year. During summer-fall 2018, the total wildfire costs in California was $24.5 billions of dollars, and it causes 106 deaths.
The spread of fire is hard to predict and control. It depends on multiple factors, such as wind, topography, humidity, complex fuel structures. Complex meteorological scenarios could cause almost instant change of the fire direction. Thus, modeling and predicting wildfires behavior requires collecting a huge amount of data in real time. This is a serious computational challenge for researchers.
Artificial Intelligence technologies, such as Big Data, Data Mining and Machine Learning (ML), could be implemented into existing physics-based models. Or they could become a basis for the new wildfires prediction systems. Such systems already exist. They use adopted Artificial Neural Networks (ANN) and Support Vector Machines (SVM) to predict the spread of wildfires. Most of them use various data sources received from meteorology stations, satellites, individual cameras. A predictive model uses satellite images and individual cameras to recognize the wildfire. Then, meteorology and topography data are used to predict the spatial spread of the fire. The result of the prediction is the 2D vector that describes fire perimeter at specified time intervals. This 2D vector can be easily visualized. So, fire managers can use it in order to plan firefighters work and evacuation of the area residents. The accuracy of such predictive systems is different and depends on ML approach, complexity of meteorological scenario, and the region specification. For example, FireCast system has an average accuracy of 87.7% and recall 91.1%.
Earthquakes
Unlike other natural disasters, earthquakes are very difficult to predict, because there is no particular mechanism of their prediction.
The earthquakes lethality and destruction rates are very high. The total number of deaths for the last decade is higher than 1.5 million. Only in the United States, earthquakes cause millions of dollars in insured losses per year.
The appearance of earthquakes are caused by movement of the tectonic plates. The crust of our planet consists of huge tectonic plates. There is a high stress in places of their connection. During the earthquake, the stress becomes critically high, the big elastic energy releases. That energy causes destruction and fault ruptures of rocks. Some part of energy transforms into vibrations, called seismic waves. The waves reach Earth’s surfaces and cause vibrations of soil, some of them are destructive. Intense ground shaking can generate landslides, surface faulting, tsunamis, floods, soil liquefaction. The earthquakes may last from seconds to minutes.
Not all of the earthquakes have natural origin, some human activity can also produce earthquakes, e.g., oil drilling, coal mining, and others.
Researchers collect huge amounts of seismic data in different regions. Although, modern models can describe earthquakes but cannot predict them. The occurrence of earthquakes seems quite random. Nowadays, the Artificial Intelligence algorithms have come to help seismologists. Machine learning methods allows researchers to detect some hidden patterns of periodicity in earthquakes occurrences. Like it was done for the Geysers geothermal field in California. Data clustering by Convolutional Neural Networks of 46,000 earthquakes in that region shows clear temporal patterns. This knowledge helps to detect changes in seismic sources.
Another way Machine Learning can help is localization of earthquakes. Thus, it was created models based on Convolutional Neural Networks to predict earthquake localization in West Bohemia and aftershocks locations in Kobe, Chi-Chi, and Kashmir. The accuracy of those models is close to 90% with low value of false positive detections (false detection of earthquake can induce panic).
Hurricanes
The 2019 Hurricane season in the USA has traditionally caused billions of dollars in losses. Hurricanes are ocean-bred storms that appear in midsummer through late fall (in the case of Atlantic hurricanes). The wind speeds in hurricanes vary from 119-153 km/h (Category 1) to >252 km/h (Category 5).
Predictions of hurricanes include two directions:
- The first is related to predictions of hurricane occurrence. Data scientists use historical data about previous hurricane seasons in order to predict the number of storms, their intensity. Statistical methods can give information about approximate wind speeds and intensity of sustained winds. Unlikely, the accuracy of these predictions is not very high.
- The second is responsible for the tracking of a current hurricane. Satellites, like GOES (the Geostationary Operational Environmental Satellite), every five minutes records images of the Hemisphere. These images of existing hurricanes give information of its activity. Then, deep learning algorithms detects the eye of a hurricane. The long-wave radar data and satellite data are combined in order to develop a predictive model. There are several widely used models: European Center for Medium Range Weather Forecast (ECMWF) model and National Weather Service’s Global Forecast System (GFS) models. Usually, predictions of hurricane path are for 3-5 days. The accuracy of such predictions decreases with time.
A Hurricane forecast is important to prepare people from dangerous regions and organize evacuation from a region with high risk.
Volcanic Eruptions
People live near volcanoes from ancient times. There are always a dozen or more volcanoes in active state. Their activity ranges from gas-and-ashes plums to lava flows. So, scientists are constantly working on methods to effectively predict volcanic eruptions.
1,500 volcanoes are constantly monitoring. The monitoring includes many factors:
- Earthquakes. Seismic data, such as number and size of earthquakes, is recorded. Since, volcano eruptions produce a sequence of earthquakes before the eruption in the vicinity of the volcano.
- Gas emissions. Sulfur dioxide (SO2) gas begins to escape when its pressure increases when magma is close to the surface. Measurements concentrations of the SO2,CO2/ SO2ratio, and hydrochloric acid gives information of volcano activity.
- Remote monitoring. Satellite technology allows us to monitor some important data about volcanoes. Satellite’s sensors give information about some volcanic gases, thermal signatures, cold eruption clouds, ground deformation, etc.
- History of volcanic activity is also important in building predictive models.
Huge amount of data and needs of real-time monitoring requires assistance of artificial intelligence technologies. An example of such assistance is the MOUNTS platform (Monitoring Unrest from Space). This platform uses multisensor satellite-based imagery (Sentinel-1 Synthetic Aperture Radar SAR, Sentinel-2 Short-Wave InfraRed SWIR, Sentinel-5P TROPOMI), ground-based seismic data (GEOFON and USGS global earthquake catalogues) and provides real-time access to these sources. Trained Artificial Neural Network detects strong deformations in the interferograms generated by the system. MOUNTS platform informs about volcano activity in real-time.
Floods
Floods are a common natural disaster. The reasons for floods are various, from rivers swollen by heavy rains to storm surges. Like other disasters, their ecological impacts can last for decades or longer.
Forecasting of flooding events requires several types of data:
- The amount of rainfall. Real-time monitoring of rainfall amounts gives information about flood risks.
- The rate of change in river or sea stage helps indicate the severity and immediateness of the disaster. Also, it is important to know how strong the stream of the river is.
- Data about a river’s drainage basin,such as soil-moisture conditions, ground temperature, snowpack, topography, vegetation cover.
- Storms path and its intensity is important for the regions close to the sea during hurricane seasons.
- Satellite’s images and aerial photography. Image processing gives knowledge about flooded areas, such as how much water runs off, where it runs to, how wide an area is flooded, and dynamic of flood.
Deep learning techniques and applications based on artificial intelligence can process different types of data, predict floods dynamic in real-time, and can keep people updated with the latest critical information about the disaster.
