In recent years, extreme weather events have been increasing worldwide. In Japan as well, serious damage has occurred across many regions. Examples include localized heavy rainfall leading to urban flooding and road inundation; landslides; damage to buildings and vehicles caused by hail; and structural damage and fallen trees due to sudden strong winds such as tornadoes. Given these conditions, there is an urgent need to establish an information infrastructure that can accurately capture what is happening in the sky right now and what is likely to happen next, and translate this information into timely and appropriate operational decisions by relevant organizations. Toshiba addresses this need by providing "Rainfall Monitoring Service" which captures the location and intensity of rainfall in near real time, "Rainfall Prediction Service" which predicts the likelihood and intensity of localized heavy rain up to 30 minutes ahead, and Japan’s first^* "Hailfall Prediction Service" which predicts hail events tens of minutes in advance and provides alerts. These services strongly support disaster management efforts as well as operational decision-making for social infrastructure operators, local governments, and private companies.

In this article, we explain the technologies for analyzing weather radar data that underpin these services, as well as the mechanisms of each service.

* Based on Toshiba research as of February 2024, referring to short-term hailfall prediction based on real time weather radar data analysis.

Weather radar is a system used to observe current atmospheric conditions. It transmits radio waves into the atmosphere and analyzes the signals reflected by raindrops (echoes) to determine rainfall intensity and location. For decades, Toshiba has provided weather radar systems to operators of critical infrastructure such as rivers, dams, railways, and power networks. These systems enable accurate observation of rainfall, gusts, and lightning. A weather radar system consists of hardware such as parabolic antennas, and software that analyzes raw radar data acquired by the antennas. While the hardware is generally common across users, the software is customized according to each operator’s specific needs. The software incorporates a wide range of weather radar data analysis technologies refined through Toshiba’s decades of experience, together with the extensive expertise the company has accumulated over the years.

By applying these technologies, Toshiba has developed a Weather Data Service that delivers highly accurate meteorological information directly to customers. Currently, this service includes three offerings: Rainfall Monitoring Service, Rainfall Prediction Service, Hailfall Prediction Service. These services are designed to make Toshiba’s advanced radar data analysis technologies accessible to a broader range of organizations, contributing to disaster risk reduction in areas such as public safety, transportation, agriculture, and event management (Fig. 1).

* For details on the Weather Data Service, including an overview, usage scenarios, and case studies, please refer to this article.

Japan operates 65 weather radars installed by the Ministry of Land, Infrastructure, Transport and Tourism (as of June 2024). ^* Since April 2022, the raw data from these radars has been made available for commercial use. Toshiba’s Weather Data Service leverages this raw data and is underpinned by the company’s proprietary technologies (Fig. 2).

* Ministry of Land, Infrastructure, Transport and Tourism website (https://www.mlit.go.jp/tec/tec_fr_000040.html) (in Japanese)

First, raw radar data originally expressed in a polar coordinate system^*1 is ingested into cloud-based analysis software to remove noise and interference. Next, the cleaned data is converted into a Cartesian coordinate system^*2 and multiple datasets are integrated to generate a three-dimensional representation of precipitation. Finally, advanced computations are performed depending on the application, such as rainfall monitoring, rainfall prediction or hailfall prediction. This processing pipeline enables both high accuracy and real time performance.

*1 A polar coordinate system here means a way of describing an observation point using distance, azimuth and elevation with the radar position as the origin.
*2 A Cartesian coordinate system here means a way of describing a observation point using three axes: east-west, north-south, and altitude.

Raw data observed by weather radars is represented in polar coordinates. Raw radar data always contains unwanted signals, such as radio interference from nearby radars and wireless systems, reflections from terrain and buildings, and instrument-specific variations. These unwanted signals distort precipitation echoes and can lead to incorrect analysis. Toshiba addresses this by detecting and removing such noise directly in the raw data stage.

For example, when radio interference from adjacent radars or various wireless sources is introduced, it may be difficult to distinguish between precipitation echoes and interference signals, which can lead to false detection of rainfall in areas where the weather is clear. To address this issue, Toshiba has successfully developed a method to detect and remove radio interference by leveraging its characteristic radial patterns centered on the radar site (Fig. 3-left).

By reducing unwanted echoes caused by non-precipitation components, we improve data quality and stability for subsequent processing, such as Cartesian coordinate transformation and rainfall prediction. These capabilities have been achieved through Toshiba’s accumulated expertise in weather radar data analysis.

Furthermore, we analyze and correct variations in data characteristics among different radar systems, enabling the generation of spatially consistent and homogeneous data over a wide area (Fig. 3-right). As a result, improvements in the accuracy of subsequent processes, such as three-dimensional compositing and hydrometeor classification, have enabled the early detection of initial signs of localized heavy rainfall.

Three-dimensional compositing is a core process that integrates observation data from multiple weather radars and represents the precipitation distribution over a wide spatial area in a Cartesian coordinate system. In observations by a single weather radar, due to the antenna beam structure, the spatial resolution^* decreases as the distance from the radar site increases, resulting in sparse observation coverage at farther ranges.

*Spatial resolution indicates how finely a space can be represented. Range resolution refers to horizontal resolution, and higher resolution enables the representation of finer terrain features and localized phenomena.

To fill these sparse regions, we employ a commonly used radar compositing technique known as the Cressman interpolation method. Cressman interpolation is a representative weighted averaging method that reconstructs a Cartesian grid (mesh) based on irregularly distributed observation data obtained from multiple weather radars (Fig. 4). For each grid point, values are calculated by applying distance-based weighting to surrounding echoes. Therefore, it is important to appropriately set the “radius of influence” according to the radar beam width and range resolution.

In operational use, the radius of influence is dynamically adjusted depending on the distance from the radar. In regions where radar observations are dense and observation points are located near the Cartesian grid points, the compositing area is kept small to avoid excessive smoothing and to prioritize high-quality data. In contrast, in regions where radar observations are sparse and no observation points are located near the grid points, the compositing area is expanded to ensure that no grid point remains unfilled. Through this approach, a gridded field suitable for estimating precipitation, snowfall, and advection vectors^* is generated.

*An advection vector represents the direction and speed at which meteorological phenomena (such as precipitation systems or wind) move in space.

The entire sequence of processing—from initial polar-coordinate processing to three-dimensional compositing described above—enables the generation of high-quality data that can be commonly used for downstream applications such as rainfall and gust prediction, as well as hydrometeor classification.

By analyzing vertically distributed precipitation data obtained through three-dimensional compositing, it is possible to estimate the total amount of water contained in the atmosphere, known as VIL (Vertically Integrated Liquid). A technology that predicts the location and intensity of rainfall several tens of minutes in advance, based on the temporal evolution and spatial distribution of VIL, is called VIL nowcasting^*.

*VIL nowcasting is a precipitation forecasting algorithm developed by the National Research Institute for Earth Science and Disaster Resilience (NIED).

Based on VIL nowcasting, Toshiba has developed its own technology that enables three-dimensional monitoring of cumulonimbus development and early signs of heavy rainfall over a nationwide analysis domain divided into 250-meter grids. Cumulonimbus clouds grow due to updrafts generated by atmospheric instability and form high-density regions of raindrops inside the clouds, known as “precipitation cores.” These precipitation cores eventually descend to the ground, causing localized heavy rainfall. Since VIL increases rapidly during the formation of precipitation cores, real time analysis of VIL allows for highly accurate prediction of localized heavy rainfall up to 30 minutes in advance (Fig. 5).

As described so far, Toshiba performs primary processing such as noise and interference removal on raw weather radar data, followed by three-dimensional compositing to analyze rainfall intensity. In rainfall estimation, different approaches are applied depending on rainfall type—light rainfall (drizzle) and heavy rainfall (intense rain). Each method has its advantages and limitations, making it challenging to achieve both wide-area coverage and high accuracy simultaneously. For light rainfall, range resolution is high, but rainfall estimation accuracy tends to be low. For heavy rainfall, rainfall estimation accuracy is high, but range resolution tends to be low.

Toshiba has addressed these challenges through dedicated technological developments. For light rainfall estimation, we developed a machine learning model trained using radar observations, terrain data, and AMeDAS rain gauge measurements. This model significantly improves the accuracy of light rainfall estimation. For heavy rainfall, we established a new analysis method that simultaneously improves both rainfall estimation accuracy and range resolution. By combining these technologies, it has become possible to estimate rainfall according to its characteristics, enabling accurate current rainfall estimation with a level of precision comparable to that of ground-based rain gauges (Fig. 6).

Weather radars deployed in Japan are dual‑polarization radars capable of observations using both horizontally and vertically polarized waves. It is possible to estimate the shape and type of hydrometeors by analyzing echoes from rain clouds observed by these radars.

Toshiba has developed a technology that utilizes parameters such as the differential reflectivity and correlation coefficient of the two polarizations to classify precipitation particles including rain, snow, graupel, and hail in the atmosphere in real time with high accuracy. For example, in hailfall prediction, precipitation particles are first identified as hail or non-hail based on raw radar data. If hail is detected, the expected landing location is forecast by incorporating the advection vector of the storm system. Furthermore, by incorporating external information such as social media (SNS) posts, Toshiba has improved the prediction success rate of hail forecasts while reducing the over-prediction rate (Fig. 7). This approach is protected by Patent No. 7532699.

Toshiba has enabled accurate, wide-area and high-resolution understanding of both “what is happening now” and “what will happen next” in the atmosphere by leveraging advanced analytics—such as noise and interference removal from raw weather radar data, three-dimensional compositing, and interpolation—together with weather forecasting technologies. These capabilities support real-time, high-precision services for rainfall monitoring, rainfall prediction, and hail prediction. In addition, we are working on the development of technologies to provide more advanced services, such as extending forecast lead times and predicting phenomena like gusts and strong winds.

Through “visualizing the sky,” Toshiba aims to improve safety, create new value from weather data, and contribute to a more resilient society.

• Weather Data Service