Multiscale storm identification and forecast

We describe a recently developed hierarchical K-Means clustering method for weather images that can be employed to identify storms at different scales. We describe an error-minimization technique to identify movement between successive frames of a sequence and we show that we can use the K-Means clusters as the minimization template. A Kalman filter is used to provide smooth estimates of velocity at a pixel through time. Using this technique in combination with the K-Means clusters, we can identify storm motion at different scales and choose different scales to forecast based on the time scale of interest. The motion estimator has been applied both to reflectivity data obtained from the National Weather Service Radar (WSR-88D) and to cloud-top infrared temperatures obtained from GOES satellites. We demonstrate results on both these sensors.