An Artificial Neural Network to Forecast Thunderstorm Location: A Search for More Relevant Land Surface Input Data

The authors are attempting to enhance the accuracy of an Artificial Neural Network (ANN) developed to forecast the location of thunderstorms. The inputs to the ANN include output from a Numerical Weather Prediction (NWP) model (numerical integration performed on a 12-km horizontal grid), and high-resolution Aerosol Optical Depth (AOD) and Land Surface Temperature (LST) observations. The AOD and LST data have a horizontal grid spacing of 1-km, thus representing sub-grid scale data with respect to the NWP output. The basic rationale for the selection of inputs is the following: The NWP output provides a forecast of the general mesoscale environment that would surround a developing convective cell. Since individual convective cells develop on the micro-scale/meso-alpha scale, data are needed that represents the forcing on these scales (not resolved by the NWP model) responsible for convection. Research has shown that land surface heterogeneity contributes to micro-scale/meso-alpha scale thermal gradients that ultimately result in convergent wind that can trigger convection. AOD may contribute to convective development by enhancing subsequent storm updraft. The ANN methodology includes a South Texas grid of 14 x 23 equidistant points that create 286 separate 20-km x 20-km box regions. A framework was established to train a separate supervised ANN for each box, to predict thunderstorm occurrence.

The goal of the present study is to provide more relevant LST-related inputs to the ANN. Specifically, to provide data that better characterizes the LST pattern responsible for generating the wind pattern that can trigger convection. The Canny edge detection algorithm was used to identify regions within each box region that represent a thermal boundary or gradients of LST sufficient for micro-scale/meso-alpha scale wind development. Two sets of results from the Canny technique were generated – one for lightning cases and the other for non-lightning cases, based on the reasoning that fundamentally divergent LST patterns would emerge between lightning and null cases. Subsequent quantification of these patterns would provide more relevant LST-related inputs into the ANN, which should result in an improvement in ANN performance. The data set represents the 2003-2006 period.

The implications of this technique are significant. Utility of this ANN model would suggest that the current paradigm of improving thunderstorm prediction by increasing the resolution of NWP models to explicitly predict convection may not be necessary. Instead, a supervised ANN can be developed that inputs mesoscale NWP model output parameters and microscale data that directly contributes to convective initiation. The utility of this method is further enhanced by studies which by inference questions the ability of high-resolution NWP models to predict convection on scales smaller than 5-km. Further, the computing resources necessary for operational NWP forecast on the micro-scale/meso-alpha scale are enormous. Yet, the ANN model described herein does not require NWP output at such high resolutions.