Weathernews AI

　Weathernews leverages IoT to collect and analyze big data, including not only weather data gathered by observation instruments, but also real-time pollen levels from pollen observation robots, wave and wind reports received from the captains of oceangoing ships, and weather reports sent in by the general public. We use this data to make forecasts and provide our Risk Communication services (countermeasure recommendations) to help our customers in various markets respond to weather risks. This service is carried out by Risk Communicators with a thorough understanding of the weather-related issues that affect each customer's business.
Incorporating AI technology appropriately in the process enables us to not only improve the accuracy and efficiency of our Risk Communication services, but in some cases even provide information and solutions that could not be achieved by man-power alone.
Weathernews calls this AI technology, LAPLACE, after "Laplace's Demon", an idea proposed by 18th century French, mathematician, physicist, and astronomer, Pierre-Simon LaPlace, which states that if we could accurately understand the present, we would be able to predict (calculate) future phenomena, which are an extension of the present.

Weathernews is using LAPLACE in various places within the company.

At Weathernews, we gather "sensory data" that reflects the human senses, from the 180,000 or more weather reports we receive daily, as well as approximately 1,000 live cameras installed around the world. Since these types of data cannot be expressed using physical quantities, such as a temperature of 18 degrees or a precipitation rate of 3 mm/h, we initially had weather experts check each one visually, but because of the enormous volume, there were more and more cases in which the contents of the reports could not be utilized in real time for the subsequent forecasting process. However, using LAPLACE we can perform natural language processing to quantify expressions within text that are unique to weather forecasts, and use image recognition to detect the formations and development of clouds from photographs. Combined with the ability to learn about feature weighting, we are now able to utilize this information in near-real time as part of the forecasting process.

The Weathernews Navigation Weather Group receives about 10,000 email messages each day from vessels currently sailing all over the world as well as from land-based operators. Each of these many important messages has its own unique characteristics and uses terms and expressive styles that are specific to the shipping industry. For example, there are words and phrases that have been abbreviated to minimize transmissions costs, originating in the days when TELEX was common, where such charges were based on the number of characters sent. In addition, each shipping company has its own email format, and the native language, culture, and personality of the person writing the email also affects the writing style. Proper training is necessary to correctly understand the context of these email messages instantly.

In the past, one person would read around 1,500 email messages per day and sort them into, "urgent items", "schedule issues", and "observations", and assign them to staff to carry out the next part of the process. However, in the timezones where the messages were concentrated, the sorting was time-consuming and it also took time to implement the necessary actions.
In 2017, LAPLACE, which had been trained to decipher these various email styles, succeeded in carrying out about 60% of this email sorting, enabling us to act more quickly on behalf of our customers.

Compared to Japan, the measurement infrastructure in developing Asian countries, such as rain gages and weather radar, is less well developed. In order to understand the actual weather situation, it is important first to develop the measurement infrastructure in the field. However this may not proceed smoothly, due to the various circumstances in each country regarding the placement of observation equipment, conditions at installation sites, etc. On the other hand, it is possible that a weather-related disaster could occur now, while waiting for that infrastructure to be completed, resulting in a situation where, even though prediction is possible, it is not possible to respond. Therefore, since the launch of the Himawari 8 satellite, which covers the Asia region, we have used deep learning to train LAPLACE to differentiate between those phenomena which correlate with the satellite images and those which do not correlate. This helps us to estimate precipitation levels in developing Asian countries and use the data as part of our forecasts.

In the future, we will introduce the use of LAPLACE into our Risk Communicators' work of informing our customers about weather risks. At Weathernews, we are now working on the development of a number of next-generation communication support tools to multiply the abilities of LAPLACE. In addition to speech recognition tools for understanding situations based on dialog with customers, these also include machine learning tools to process various types of meteorological data, tools to generate countermeasure information that is customized to the characteristics of each customer, and content matching technology.

For example, even a small amount of rainfall during asphalt replacement for road repairs can diminish the quality of the finished surface, requiring that the work be redone, which leads to increased costs for process readjustments, etc. Even with the latest technology, forecast accuracy for phenomena that cause such problems is not 100%, but losses can be reduced through good detection of the circumstances at the time and appropriate countermeasures. Therefore, our system reads in and analyzes large scale data on moment-to-moment and changing weather phenomena, along with data extracted from customer dialog regarding things like, "risks and events of concern", "the customer's circumstances", "degree of urgency", and then generates information such as, "the level of risk the customer faces" and "what countermeasures can be taken at this time".

By communicating with many customers through the operation of this system, we learn about what kinds of information interests our customers, under which circumstances, as well as related trends, and use this information to improve the quality of our support services.