Exploring Doppler Weather Radar Images to forecast local area rainfall
Day to Day weather change plays an important role in agricultural production. Accurate prediction of rainfall significantly influences crop growth, development, and yields, pests and diseases; water needs; and fertiliser requirements. Erratic weather is beyond human control. However, it is possible to adapt to or mitigate the effects of adverse weather if a forecast of the expected weather can be obtained in right time.
Doppler weather radars (DWR) are used to locate the position and velocity or motion of precipitating systems like rainfall/snowfall to nowcast the storm events. They send radio waves in the atmosphere which when interact with precipitating systems reflect back and give information about their position, velocity as well as the type of hydrometeor present in them. These are extensively used in assimilating real-time precipitation data in Numerical Weather Prediction models to nowcast the weather.
Nowcasting means predicting the state of the atmosphere just 1–6 hours ahead of time. Radar nowcasting utilizes the vector motion of precipitating fields to extrapolate them in future time.
In this article, I am sharing our experience of exploring the potential of Radar/Satellite Rainfall Imagery to identify the rainfall clusters and locating them on geo-coordinate systems to generate rainfall alerts for nearby locations. Pillow and Scikit-Image are two of the very important libraries being used in Image processing.
Pillow can be installed via command:
pip install PillowBelow the figure is a typical DWR map of surface rainfall intensity where the centre of the image represents the Radar location and concentric circles embedded in the image represent the radial distance from the centre. Blue patches are the rain-bearing pixels.
Typical radar image update frequency is 10 minutes for the Indian Meteorological department. i.e. every 10 mins a new image is updated with the latest state of the precipitating system. The scikit-image processing module can be utilised in identifying the rainfall clusters and their morphological features like shape, size, orientation, eccentricity, etc.
Once rainfall clusters are identified, they can be mapped on the geo-coordinate system to calculate their approximate distance from any location. However, when the system/cluster moves, its motion vector plays an important role in nowcasting or extrapolating the future position of the cluster with time.
This method follows the idea to navigate the rainfall clusters and find how far they are from any user location. It needs only real-time images but a further good future scope of work lies in nowcasting, the storm, or extrapolating the exact motion, time, and intensity of rainfall cluster for a 1–6 hour time horizon using deep neural nets. However, while nowcasting the rainfall through Radar Images, it is important to keep a check on whether the rainfall cluster is getting intensified or de-intensifying with time.
A major limitation of using radar images in nowcasting the rainfall is its maximum range. Radar gives you a finer picture of rainfall but on a smaller spatial scale i.e. 80–240km range. So, Nowcasting the intensity and motion of larger or bigger rainfall systems that have their spatial coverage more than radar range become tedious.
