According to a UN report, more than 50% of the total world's population resides in urban areas and this fraction is increasing. Urbanization has a wide range of potential environmental impacts, including those related to the dispersion of potentially dangerous substances emitted from activities such as combustion, industrial processing or from deliberate harmful releases. This research is primarily focused on the investigation of various factors which contribute to the dispersion of certain classes of materials in a complex urban environment and improving both of the fundamental components of a fast response dispersion modeling system—wind modeling and dispersion modeling. Specifically, new empirical parameterizations have been suggested for an existing fast response wind model for street canyon flow fields. These new parameterizations are shown to produce more favorable results when compared with the experimental data. It is also demonstrated that the use of Graphics Processing Unit (GPU) technology can enhance the efficiency of an urban Lagrangian dispersion model and can achieve near real-time particle advection. The GPU also enables real-time visualizations which can be used for creating virtual urban environments to aid emergency responders. The dispersion model based on the GPU architecture relies on the so-called "simplified Langevin equations (SLEs)" for particle advection. The full or generalized form of the Langevin equations (GLEs) is known for its stiffness which tends to generate unstable modes in particle trajectory, where a particle may travel significant distances in a small time step. A fractional step methodology has been used to implement the GLEs into an existing Lagrangian random walk model to partially circumvent the stiffness associated with the GLEs. Dispersion estimates from the GLEs-based model have been compared with the SLEs-based model and available wind tunnel data. The GLEs-based model is more dispersive than the SLEs-based model in both the lateral and vertical directions. It is observed that for the present test case, the GLEs-based model performed relatively better than the SLEs-based model.