A feed-forward, supervised, multi-layer perceptron Artificial Neural Network (ANN) was developed to forecast the initial occurrence of thunderstorms 9 hours in advance, within 20-km x 20-km horizontal regions in South Texas. The authors are testing the hypothesis that an ANN can be developed to successfully forecast thunderstorm occurrence within a 400 km2 region by providing two sets of inputs into the foregoing ANN. The first set consists of sixteen (16) output parameters extracted/derived from a hydrostatic numerical atmospheric model (Eta), wherein numerical integration was performed on a North American grid with 12-km grid spacing. The second set consists of sub-grid scale (1-km grid spacing) Land Surface Temperature (LST) data derived from the Moderate Resolution Imaging Spectroradiometer (MODIS) output. By definition, the Eta cannot explicitly account for this sub-grid scale data. Yet, LST gradients contribute to meso-scale atmospheric motions that can initiate thunderstorm activity. For the thunderstorm target, cloud-to-ground lightning strike data from the National Lightning Detection Network (NLDN) was used. This strategy represents a departure from the current paradigm of decreasing the horizontal grid-spacing (which is extremely computer intensive) to approximately 2-km to explicitly forecast convection. ANN performance was evaluated for two 400 km2 regions in South Texas: A coastal setting, which includes the city of Corpus Christi, and an inland region further north which includes the city of Victoria. The performance was evaluated by comparing actual to predicted lightning, using several parameters including the Probability of Detection (POD), False Alarm Rate (FAR), and the Heidke Skill Score. In addition to the results, the relative importance of the input parameters will be discussed as well as the future operational applicability of the model and differences between the two selected study locations.