Local precipitation nowcasting using artificial intelligence using weather radar maps

Local precipitation nowcasting using artificial intelligence using weather radar maps

Whether it's walking the dog or a morning commute – an up-to-date and accurate weather forecast can mean the difference between a pleasant walk under an umbrella or getting wet to the bone.

Current weather forecasts are based on traditional weather forecasting (NWP) models that use complex mathematical algorithms representing physical atmospheric principles to predict how the weather will change over time. A variety of real-time meteorological observations are generated from ground-based sensors, weather radars, and satellites that monitor land masses, oceans, and upper atmosphere. Variables such as humidity, temperature, wind direction and speed, etc. are fed into the models to produce daily and long-range forecasts for up to two weeks ahead at regional, national and global levels.

However, such traditional methods can face difficulties when it comes to short-term and small-scale precipitation forecasting, due to incomplete observational data, incomplete modeling, or insufficient understanding of the atmosphere.

To provide accurate precipitation forecasting, Beijing-based startup ColorfulClouds Tech applies machine learning (ML) techniques using observed radar echo maps to generate high-resolution, minute-by-minute precipitation forecasts on its MobileClouds Weather app.

Weather forecast.jpg
Sample minute-by-minute precipitation forecast in the ColorfulClouds Weather app

ColorfulClouds has created a comprehensive model that takes reliable radar mapping data from China's NMIC (National Meteorological Information Center) and the US National Oceanic and Atmospheric Administration (NOAA) to complete forecasts through technical processes that support machine learning for segmentation and forecasting.

Segmentation filters and removes false echoes from non-precipitation objects such as buildings, hills, birds, and airplanes that may appear in radar reflectivity images. The model training is based on U-Net and SegNet. The segmented images are then processed using deep neural networks to create time-bound fuzzy predictions and output future radar maps.

False echo.jpg
Echo signals from Beijing's buildings and mountains were reflected in the radar image
Before (left) versus (right) segmentation of the radar map image

The company uses a self-developed tool to classify data. In China, this data covers maps from more than 200 radar sites for different seasons, latitudes and terrain to ensure a high-quality dataset for model training.

ColorfulClouds says the weather forecast accuracy of its latest ML models for the next six hours is comparable to traditional weather forecasting models.

The ColorfulClouds Weather app is designed to complement traditional weather forecasting methods, allowing users to stay informed of local weather conditions such as temperature, pressure, and wind; Providing additional information for public awareness (such as smog). Air pollutant concentrations and Air Quality Index (AQI) data are currently available for Asia and North America.

aqi demo.jpg
Sample air quality data from ColorfulClouds Weather

Besides nowcasting of precipitation, ColorfulClouds is also considering applying machine learning to other aspects of weather forecasting, such as extending accurate short-term precipitation forecasts to longer-term precipitation forecasts.

Machine learning is not new to weather forecasting. Meteorologists and researchers are constantly developing and improving data assimilation and forecasting algorithms and updating relevant parameters. Common machine learning applications in contemporary weather forecasting include tropical cyclone intensity forecasting and extreme convective weather forecasts.

ColorfulClouds believes that machine learning is also applicable in numerical weather prediction (NWP) models that require fuzzy estimation, data interpolation based on observation, parameter estimation, data assimilation, etc. However, there is still a talent shortage, as very few professionals in this field have a comprehensive understanding of meteorological models.

Inspired by NeurIPS Best Paper 2018 Neural ordinary differential equationsColorfulClouds' long-term goal is to develop deep neural networks that allow backpropagation of the set of equations used in weather forecasting. Discovering a machine learning approach that outperforms NWP models that have been tried and tested for 40 years will be both challenging and rewarding.

This article is based on a concurrent interview with the manager of the ColorfulClouds algorithm team Pengqiu Xu. Founded in 2014, ColorfulClouds provides weather forecasting and translation services.

source: China Concurrent

the site: Tingting Cao | editor: Michael Sarrazin

You may also like...

Leave a Reply

Your email address will not be published. Required fields are marked *