InterMET Session 2: AI & machine learning
15:30 - 16:30
CHAIR: Jim Anderson, Chairman, Association of Hydro-Meteorological Equipment Industry
Developments in the use of AI & machine learning in the interpretation and application of weather and climate data.
- Dr Amy McGovern, University of Oklahoma
Title: What is trustworthy AI when it comes to the prediction and understanding of high impact weather events, and why do we need it?
- Dr Victoria Bennett, Head of User Services, European Centre for Medium-Range Weather Forecasts
Title: ECMWF strategy on the role and use of AI and machine learning in provision of weather-related data products
Dr Julien Oliver, Brand Technical Specialist, Climate, Environment and Sustainability, IBM AI Applications - ASEANZK
Title: Democratising Climate Risk and Impact Analytics
Abstract: As the United Nation’s Intergovernmental Panel on Climate Change report blatantly states, this is just the beginning. Temperatures will likely continue to rise triggering more frequent and intense extreme weather events. We have to continue our efforts to slow down climate change. But we also have to adapt to our rapidly changing environment.
Organizations often lack access to specialized tools and data geared towards expediting and improving climate risk and impact analytics. Over the past years, IBM has developed unique frameworks and technologies integrating AI with environmental, climate, and weather data to help companies predict, prepare for and adapt to the increasingly severe risks of our warming climate.
IBM’s environmental and Climate Impact Modeling Framework, CIMF, relies on the capabilities of IBM Environmental Intelligence Suite (EIS) and PAIRS++. The platform addresses the most challenging barriers organizations face today when attempting to model climate risks, including:
- Opening up access to large-scale computing power through the cloud.
- Streamlining data inefficiencies when collecting unstructured, heterogeneous data (aerial imagery, maps, IoT infrastructure, drones, LiDAR, and satellites).
- Standardizing weather forecasting models (e.g. The Weather Company, ECMWF) into an accessible and easily interpretable framework
In this presentation, we will introduce the concept and science behind IBM’s CIMF and illustrate with regional case studies how organizations can efficiently monitor or produce probabilistic flood risk analysis from historical climatology or based on seasonal weather predictions.
- Brian Bellew, Global Business Development Manager (Pacific, South Asia, East Africa), Baron Weather
Title: Machine Learning Addresses Challenges Resulting from Increasing Radio Interference to Improve Radar Data
Abstract: Weather Radars produce images of hydrometeors or other objects in the atmosphere and are one of the primary meteorological tools used by scientists, forecasters, and researchers worldwide. The recent expansion of RF communication and other technologies or even naturally occuring phenomena has resulted in interference and distortion within the radar signal. A weather radar uses reflected radio waves to locate and create images of precipitation and moisture. To be correctly interpreted operators require training, however, it is standard practice in many Hydrometeorological agencies around the world to push these images to the public for consumption. The interference on the radar produces false images that might be interpreted as precipitation making it difficult for the novice user to understand what they see. Also, many short-term weather models use radar data to predict precipitation development, and thus the highest quality data is important for these predictions. Baron has studied the problem extensively and found that using machine learning, the radar data can be “cleaned.” We will demonstrate how advanced machine learning technology has improved radar imagery and discuss the benefits and importance of clarity in the generation of weather radar imagery.
Summary: The invention and application of the weather radar was one innovation that dramatically changed the science of meteorology and aided in the creation of weather forecasts. Even the general public routinely uses weather radar data, and in many places around the world, it is the most popular digital weather tool available today. But as all meteorologists know, weather radar is not perfect. Understanding the images, the radar creates requires training and experience, this is especially true because the radar produces images that contain echoes that are not precipitation or moisture. A weather radar operates using one of three radio frequency bands, X, C, or S. Many other electronic devices (if not controlled) can use these frequencies as well, confusing a weather radar as it listens for its own radio waves to return. In addition, other things such as wind turbines or certain atmospheric phenomena can cause radio waves to return to the radar and create images that appear to be precipitation.
Baron Weather scientists have been at the forefront of minimizing the effects of ground clutter and anomalous propagation with the CLEAN-AP™ filtering process. To combat the increasing interference from these new sources, Baron scientists turned to machine learning to solve the challenge. Baron Weather has developed a solution to “clean” a radar image. The result is a dramatically improved radar image. Once the process “learns” what is important and what is not, the machine learning tool automatically processes the images without any human intervention. Machine learning is executed at a point in the process that if the operator desire to look at the uncleaned image that is possible. The images can even be displayed side-by-side allowing the operator to study the effects of the process and build trust in the system.
Why is this important? First, a clean image can shorten the time it takes a meteorologist to make a time-sensitive decision. Knowing they are looking at only real precipitation allows the forecaster to focus on their predictive decision and less on determining what is real and what is not. Second, many shorter-term weather predictive models collect radar data to aid in the model forecast. If these models are initialized using false echoes that are not real, they can produce poor forecast output. Clean images aid the model in understanding the dynamics of real precipitation occurring over the model’s domain. Finally, as mentioned earlier radar images are very popular digital weather elements displayed to the general public. Providing them with clean images makes the radar a more useful tool to the general public and thus can increase usage.