JP6.19
A robust neural network ensemble approach for snowfall detection using AMSU data
Yajaira Mejia, NOAA/CREST at City College of New York, New York, NY; and H. Ghedira and D. R. Khanbilvardi
Precipitation is a key factor required for hydrological, hydro-climatological and meteorological models applied for flood and weather forecasting, climate change prediction, water resource management and many other applications. Throughout the years snowfall timing and intensity have had an important effect on human lives. Snowfall affects everybody directly or indirectly. Most importantly, the extent of any specific storm is undoubtedly vital information for public safety and service management. Having an accurate detection of snowfall is indispensable for an efficient meteorological forecasting and modeling. Passive microwave remote sensing techniques have been investigated by numerous researchers using various sensors and were found to be potentially effective for detecting and estimating snowfall. Most of these researches are underpinned by the hypothesis that brightness temperatures measured on high frequencies (greater than 89 GHz) are dependent on physical properties of the snow and rain particles with different correlation levels.
The principal intent of this research is to investigate the potential of passive microwave data from AMSU in detecting snowfall and evaluate the effect of both land cover and atmospheric conditions on the retrieval accuracy. A neural-network system has been developed and has shown a great potential in detecting snowfall events. This algorithm has been applied for different snow storms occurred during three winter seasons (2005-2007) in the North-East of United States. Additional information such as cloud cover, wind speed and air temperature were added to the process to reduce misidentified snowfall pixels. Only pixels with cloud cover and falling within a specific range of temperature and wind speed are presented to the snowfall detection model. Surface temperature and wind speed data collected from ground station-based observations and archived by the National Climatic Data Center (NCDC) were used for this test. Different heavy storm events and non-snowfall observations that occurred at the same time as AMSU acquisition were selected. Snowfall information retrieved from radar data also collected by the NCDC was used as truth data to train and validate the model. A minimum three hour storm was chosen to compare snowfall pixels retrieved from radar with satellite data to reduce the risk of erroneous identification of snowfall pixels used as truth data. The results indicate that the neural network-based model provides a significant improvement in snowfall detection accuracy over existing satellite-based methods.
Joint Poster Session 6, Improvements to NWP and Short-term Forecasting
Wednesday, 14 January 2009, 2:30 PM-4:00 PM, Hall 5
Previous paper Next paper