Machine Learning and Meteorology: What Does the Future of Forecasting Hold (Credit: American Meteorological Society and American Geophysical Union)

Time and time again, the world of meteorology continues to find new avenues to explore in terms of representing the state of the Earth as we know it. The advent of more powerful supercomputers and more advanced algorithms give scientists more information and finer precision on details such as surface temperature, wind speed at various altitudes, and rainfall accumulations among others. It is also the plethora of available data that makes it increasingly challenging for forecasters to monitor all available data all of the time. So what if algorithms could be created such that they can attempt to predict future states of the atmosphere based on already recognized patterns of cloud types or storm shapes? Machine learning can help to accomplish this task. Machine learning is a technique in which the model can effectively attempt to guess future outcomes based on spatial pattern recognition from previous data (analogs) and then “learn” and adapt the forecasts with the purpose of enhancing forecast accuracy and increased likelihood of forecast verification.  
 
A very recent article accepted for publication by Weyn et al. (2019) in the American Geophysical Union’s Journal of Advances of Modeling Earth Systems broaches the subject of predictability by virtue of a concept known as a convolutional neural network (CNN). Here, the authors construct a model in which past meteorological data is used to “train” the model to then predict mid-level atmospheric heights with no knowledge of ongoing physical processes. The experiment consisted of several schemes The results are optimistic in that the CNN is able to outperform basic climatology and persistence forecasts with a lower root-mean-squared error for lead times between 3 and up to 14 days from forecast initialization. Times beyond the short-term do introduce sensitivities into the model that may lead to degraded physical representations of the atmosphere as the authors note, but present-day forecasting techniques are also subject to possibly larger errors with greater lead time.
 
Now, most of the machine learning/artificial intelligence practices are still in the experimental phase as there is still much to learn about how to apply the techniques to real-time forecasts. As there are multiple components that feed into a forecast such as data assimilation, model physics, etc., present-day forecasting still has quite the ways to go before becoming a feasible option for daily use in the field. In addition, machine learning can be a cost and computer resource-intensive approach that may not be the most ideal approach given current resources. However, machine learning techniques are already making inroads in the retrospective analysis realm (reprocessing previous events and their forecasts). These retrospective analyses could become the building blocks for where to go in the near future in advancing machine learning approaches for weather forecasting. Applications of retrospective forecasting of small-scale events is discussed in a relatively recent paper by Gagne et al. (2014) for which they applied machine learning techniques to increase reliability of precipitation forecasts. Even so, machine learning also has potential applications to climate forecasting as the variability scale for climate versus weather is much smaller in time.
 
Nevertheless, artificial intelligence via machine learning is an increasingly considered tool in weather and climate forecasting and while there are still several factors that are limiting a full-scale rollout of these techniques into operations, the benefits of recent research and promising results means that machine learning could ultimately become a useful tool in the arsenal for forecasters and researchers alike.
 
Here are the appropriate citations for each article above:
 
Gagne, D. J., A. McGovern, and M. Xue, 2014: Machine Learning Enhancement of Storm-Scale Ensemble Probabilistic Quantitative Precipitation Forecasts. Wea. Forecasting, 29, 1024-1043, https://doi.org/10.1175/WAF-D-13-00108.1
 
Weyn, J. A., D. R. Durran, and R. Caruana, 2019: Can machines learn to predict weather? Using deep learning to predict gridded 500-hPa geopotential height from historical weather data. J. Adv. Model. Earth Sy,. Accepted., https://doi.org/10.1029/2019MS001705
 
Image credit above: NCAR Research and Applications Lab

​To learn more about other weather research topics and issues, be sure to click here!
 
© 2019 Meteorologist Brian Matilla