How Your Smartphone Could Help Predict The Weather

By Conor McNicholas

If you were in Seattle on Monday, February 4th you awoke to a surprisingly rare sight. Up to 8 inches of snow blanketed the Puget Sound lowlands. This was in stark contrast to forecasts on Sunday night, which called for a mere 1–2 inches of snow in the Seattle area. For meteorologists like me, Monday was a humble reminder of the challenges facing modern weather prediction.

State-of-the-art weather models, which simulate the physics of our atmosphere, still struggle to accurately forecast when and where severe weather will occur. From heavy bands of snow over Puget Sound to tornadic thunderstorms in the Great Plains, our models have difficulty pinpointing the location, intensity, and timing of highly local weather phenomena. My research aims to help solve this problem by using millions of smartphones measuring atmospheric pressure to better inform weather models.

Recent increases in computing power have facilitated rapid advancements in the resolution, or level of detail, our weather models are able to depict.

 
engage_weather1.png
 

This is exemplified in the two images above, which show simulated radar images from a decade ago (left) and from a model run today (right). While improvements in model resolution have allowed us to create more realistic looking storms, more resolution hasn’t resulted in more accurate forecasts. A good forecast is precise and accurate.

 
engage_weather2.png
 

Our high-resolution models are precise; they can depict and simulate storms with great detail, but they are not always accurate. A recent study found that even with increased model resolution, thunderstorm forecasts were off by an average of 60 miles (Duda and Gallus, 2013). The authors of the study concluded improving forecast accuracy requires more observations, not additional model resolution.

Observations are critical because they tell our models where to begin. If your model can’t accurately simulate the present, it won’t be able to predict the future. As model resolution increases, more observations are needed to inform the model.

This is where you come in. Chances are, your smartphone is one of a billion that contain a barometer. Our atmosphere has weight and it presses on us from every direction. Barometers measure this pressure. Since pressure decreases rapidly with height, pressure measurements are typically reduced to sea level. To perform this reduction, the elevation of the smartphone must be known. Estimating elevation using GPS is not feasible, as it fails indoors and is notoriously inaccurate. Instead, we assume all smartphones are at ground level. This imperfect assumption results in large errors in smartphone pressure measurements. Fortunately, these errors can be corrected using machine learning. During this process, a computer learns how sensor and GPS data relates to smartphone pressure errors. Once these relationships are known, sensor and GPS data from each smartphone can be used to correct smartphone pressure errors.

 
engage_weather3.png
 

Visualizing this process, the two images above show uncorrected (left) and corrected (right) smartphone pressures from an app I developed, uWx. Before correction, smartphone pressures varied unrealistically. In the atmosphere, pressure generally varies smoothly. Corrected smartphone pressures better reflect this reality. Remarkably, after correction, you can see a gradual change in pressure, from high to low, across Seattle, from south to north.

By providing an unprecedented level of detail, smartphone pressures may help advance our understanding of complex problems like thunderstorm formation. Smartphone pressures may also be used to inform and potentially improve the accuracy of our weather models. The future of my research will focus on testing these two hypotheses. Depending on the results, your smartphone may one day contribute to more accurate snow forecasts, preventing future snowy surprises.