UMBC ebiquity
American Geophysical Union Fall Meeting Abstracts
Using Lidar and Machine Learning to Identify Planetary Boundary Layer Heights
December 1, 2019
Accurate reporting of Planetary Boundary Layer (PBL) height is critical for weather prediction, pollution studies, and air quality monitoring. The majority of aerosols are contained in the PBL, therefore accurate measurements of its height can result in a more accurate prediction of the volume for aerosol dispersion. For general circulation models in particular, accounting for the influence of the boundary layer, which has atmospheric dynamic and thermodynamic effects, has a direct impact on the accuracy of the prediction model. As reported by the National Oceanic and Atmospheric Administration, there are more than 900 Automated Surface Observing System (ASOS) units in the United States collecting data pertaining to atmospheric data that includes Lidar data. Ceilometers are used to gather cloud base heights and Lidar backscatter. Typical techniques such as the gradient method, curve fitting, and the wavelet covariance transform method are used to estimate PBL heights from Lidar backscatter information. These methods can be problematic when residual layers are present or when cloudy conditions exist, resulting in discrepancies in estimations when compared with radiosonde measurements.
A joint research effort between the Joint Center for Earth Systems Technology and the University of Maryland Baltimore County (UMBC), funded by the National Aeronautics and Space Administration (NASA), is underway to show how machine learning could be used to automatically identify PBL heights using Lidar backscatter profiles. By processing backscatter profiles over time intervals as imagery, limited preliminary results using a machine learning approach, namely a deep neural network to identify edges in the image, have shown the machine learning method outperforms a double derivative method that included signal denoising, feature enhancements, and statistical averaging of consecutive mixing layer heights. Results from this study will provide a quantitative comparison between PBL heights obtained from radiosondes, a traditional wavelet method, and a method resulting from this research effort which entails deep transfer learning and edge detection.
InProceedings
American Geophysical Union