Human Forecasters and Technology

Harold Brooks
NOAA/ERL/National Severe Storms Laboratory
Norman, Oklahoma

(In preparation for a presentation on Challenges for Technology Developers to be made at the 5th AES/CMOS Workshop on Operational Meteorology, 28 February-3 March 1995, Edmonton, Alberta and as part of an ongoing effort to define the future state of weather forecasting.)


The relationship between human forecasters and technology is a complex one. I'd like to offer some observations about the relationship to serve as a foundation for considering what the relationship should be in the future. Comments are, of course, welcome. My e-mail address is As always, note that the following are personal opinions and do not necessarily reflect the views of any individual or organzation up from me in the chain of command (NSSL, ERL, NOAA, DoC, etc.).

Many of the thoughts expressed here are not original. I've tried to reference those things which I know have appeared elsewhere or which illustrate an idea particularly well. However, I've had discussions with many people on these and related topics and I've read quite a bit over the years. The papers by Tennekes[1] and McIntyre[2] have been particularly thought-provoking. I've had many illuminating discussions with Chuck Doswell and the presentation on the future of forecasting by Mike Fritsch at the 14th AMS Conference on Weather Analysis and Forecasting have also had significant influence on me. It is quite possible that I've assimilated some of the notions of others without remembering the original source. If there is something that should be referenced, but is not, please let me know.

Historical Perspective

In a sense, weather forecasting is impossible without a fairly sophisticated level of technological development. The development of sensors to detect the present state of the atmosphere and a communication system to collect and disseminate the data rapidly are essential. From that perspective, weather forecasting has only truly been possible from some point in the first half of this century. The old statement that "Never, no matter what the development of science, will men, who have concern for their reputation, attempt for forecast the weather"[3] is partially rooted in the 19th century difficulty of knowing enough about the current state of the atmosphere.

The ability to make and distribute observations has always been viewed as a benefit to human forecasters and it is hard to imagine any way in which anyone could view it as a threat to the role of human forecasters. A second technological development, electronic computing, has been seen both as a benefit and as a threat. The idea of numerical forecasting of the state of the atmosphere was presented in the 1920s by L. F. Richardson[4], but the technology to carry out such an effort was not developed until World War II. In the years immediately following the war, early electronic computers were applied to a number of problems, including numerical weather prediction (NWP). Early models were physically very simple and had coarse horizontal and vertical resolution. As a result, they could not capture most of the structures revelant to forecasting sensible weather.

Over the years, however, has computing power has increased and as research using the observations has improved our understanding of the atmosphere, the models have become more sophisticated and resolution has improved. Improved predictions of the fluid dynamical state of the atmosphere, coupled with statistical analysis of the model and observed atmosphere, have improved NWP forecasts.[5] Numerical "guidance" has grown from being something to which forecasters would pay little attention to the point that it dominates the forecasting process for many people. Indeed, one of the problems that many forecasters have is that they spend a lot of their forecast time trying to decide between conflicting sources of guidance, without having any scientific basis for making the decision on a given day.

The accuracy of NWP guidance has reached the point that, in many cases, there is little that human forecasters can do to improve upon it. In this regard, NWP is seen by many as a threat to the existence of human forecasters. Rather than being an ally, as in the case of making observations, technology is viewed as an "enemy" of humans involved in the forecast process. There is no reason to believe that the quality of NWP will not continue to improve and, perhaps more importantly, the lack of opportunities to add significant value to NWP guidance leads some to envision a future in which forecasts are made without human intervention.

Traits of machines and humans in the forecast process

What are the traits that help define a proper role for the machines and humans in forecasting? Machines are uniquely suited to doing repetitive tasks or tasks in which it is either unsafe or extremely costly to use human beings. As examples, I've already mentioned their strengths in observations and communications. Clearly, reducing a huge number of individual radar returns down to a single image on a screen is something that requires machines. Further, the ability to carry out large numbers of computations in a very short time makes them the only avenue for numerical solution of the equations of motion in NWP.

Human beings possess skills which are useful for forecasting (which is the main point of the McIntyre reference), particularly in the processing of visual data. Pattern recognition is not a simple task to program, but it is one at which human beings are particularly adept. Consider as an example, the recognition of a "chair". As I type this, I use a chair with no back which requires me to kneel on an angled platform. It looks nothing like the other chairs in my office--one with four legs and the another with four wheels radiating out from a central post. Yet, even my three-year-old daughter recognizes each as a chair.

As has been pointed out by many others (e.g., Pettersen[6], McIntyre, and Carr McLeod at the 13th AMS Conference on Weather Analysis and Forecasting), the essential question as to whether a machine should do a job instead boils down to using the machine to allow the person to do things people do well. In other words, using a machine to plot a large number of observations is an excellent use of machine resources because of the repetitive nature of the task. Having a machine analyze the field represented by the observations is not a simple question. While the objective analysis may capture the gross features of the field, there are numerous benefits to human analysis--recognition of "unusual" observations, which may represent errors or significant changes that indicate that an important weather event is imminent. The answer to the question of subjective versus objective analysis will not always be the same, depending upon the reason the analysis is being carried out.

At the other extreme, reacting in a severe weather warning situation is most efficiently done by human beings. While the processing of radar data should be done in an automated fashion, it has been my experience that there are human beings who are sufficiently skilled in pattern recognition and the application of the appropriate conceptual model to outperform radar algorithms for severe weather. In this case, the machine can do the routine, "mundane" tasks, while the human is doing something to which humans are uniquely suited.


Based upon the question of wheher a human or a machine should do a particular weather forecasting job, what does the future hold? I believe that "routine" forecasts, such as temperature and probability of precipitation will become, in general, more the domain of machines instead of human forecasters. The value added by humans to those machine forecasts is getting less and less, in large part due to the improvement of the machine products. It is not clear that the potential added value of significant human effort will make that effort worthwhile on most days for the currently-sized forecast areas of the National Weather Service. In 1993, Model Output Statistics (MOS) from the Nested-Grid Model (NGM) missed second period temperature forecasts by more than 5 F 15% of the time and by more than 10 F 2% of the time. Using the latter as a gross error, which humans could, at least in theory, improve significantly upon, one finds that it takes a forecast area equal to about 35 MOS sites before a forecaster, on average, has a 50% chance of making a significant improvement in temperature forecasts in any given forecast cycle. This represents a much larger forecast area than currently exists. Lance Bosart (SUNY-Albany) told me the results of a study he had carried out involving the national forecasting contest that found that proximity to a forecast location does not improve a forecaster's skill. Thus, the argument that there is a need for a large number of forecast offices in order to ensure that forecasters are "in the vicinity" is not valid.

A case can also be made that increasing technology means that it is not even necessary that forecasters be in the immediate area to make good warning decisions. Communication from storm spotters and from radar data can now be made over great distances at an imperceptible loss in information. While I am undecided on this question, there is great merit to the argument that the NWS can get by with significantly fewer warning offices and, directly related to that, with a smaller number of better-trained, high-quality warning meteorologists. A strong case can be made that the public radar warning system can be maintained at at least the current level of performance with perhaps only 10 offices with increased staff over current offices.

The improvement in technology over the decades should allow humans to concentrate on extreme events, be they gross model errors or life and property-threatening weather. Technology, which initially allowed humans to make routine[7] weather forecasts, will soon close that avenue of human endeavor, leaving open concentration on severe events. It is likely that those of us working in meteorology in the second half of the 20th century have seen a kind of human involvement in forecasting that will be open only briefly.