Near Real{time Data Assimilation for the HYSPLIT Aerosol Dispersion Model

Konstantinos Kalpakis, Shiming Yang, and Yaacov Yesha Department of Computer Science and Electrical Engineering University of Maryland Baltimore County 1000 Hilltop Circle, Baltimore, MD, U.S.A. {kalpakis, shiming1, yayeshag}@csee.umbc.edu ABSTRACT We are working on an IBM-funded project seeking to develop a prototype system for real-time plume dispersion and fire and smoke detection and monitoring. Our prototype system utilizes HYSPLIT and observation data from various sources. HYSPLIT is a model developed by NOAA's Air Resources Laboratory for forecasting aerosol trajectories, dispersion, and concentration from emission sources. It is used extensively by NOAA to routinely provide a number of data products. We develop a data assimilation system for assimilating observational data into the forecasting model in order to improve its forecasting accuracy. Our system is based on the Local Ensemble Transform Kalman Filter (LETKF) algorithm and it is computationally efficient. We evaluate our data assimilation system with real in-situ observational data, and find that our system improves upon HYSPLIT's forecast by reducing the normalized mean squared error and the bias. We are also experimenting with assimilating MODIS data with HYSPLIT model forecasts. To this end, we extrapolate ground concentrations from MODIS Aerosol Optical Depth (AOD) data. Our extrapolation approach relies on spatially localized linear regressions of aerosol concentrations from ground stations in the Air Quality System (AQS) network and MODIS AOD data. We expect that assimilating the extrapolated concentrations leads into further improvements of HYSPLIT forecasts. Furthermore, we are investigating using additional sources of in-situ and remotely sensed observations, such as GOES AOD 30-minute data, and UAV data from the Ikhana AMS fire missions. These sources provide higher spatial resolution and more frequent temporal coverage. Moreover, GOES and UAVs provide near-real time data which should be useful in improving HYSPLIT forecasts of smoke from wildfires. Currently, the Ikhana AMS fire missions team provides L1B data which are very useful in themselves, but no level 2 to the best of our knowledge. For our application, it would very useful to have an AOD data product for these datasets. A possible path for deriving AOD data the AMS sensor onboard UAVs would be to utilize the DRL code for deriving the MODIS AOD from MODIS L1B data, due to the sensor similarities. Developing such code would be very useful for wildfire smoke prediction applications. Our near real-time data assimilation system helps in bridging the gap between predictions and real-time observations, for more accurate and timely aerosol dispersion forecasts. Keywords: data assimilation, HYSPLIT, forecast model performance, real-time, ensemble Kalman filter, aerosol dispersion and concentration.