Using Machine Learning to Identify Planetary Boundary Layer Heights for Ceilometer-Based Aerosol Backscatter Retrievals

Abstract:

The planetary boundary layer height (PBLH), used to determine vertical mixture of air pollutants is critical for air pollution assessments but at times can result in inaccurate calculation due to uncertain conditions. In particular, nighttime collapse of the PBLH near the surface presents a challenge for model forecasts. Ceilometer-based aerosol backscatter profiles and gradient-based methods can be used to derive heights for station-specific geographical locations. A research effort is underway at the University of Maryland Baltimore County (UMBC) which will explore how the latest advancements in machine learning research can be utilized to automatically derive PBLH from ceilometer backscatter using deep learning. This research effort includes evaluating how a deep neural network edge detector for boundary height detection can be used to automatically identify Mixing Layer Height (MLH) using backscatter profiles from multiple stations at different geographical locations. We have been able to show the performance of the deep boundary layer height detection method when applied to ceilometer aerosol backscatter profiles acquired during the Ad-hoc Ceilometer Evaluation Study (ACES) at the UMBC for the period of December 1- December 15, 2016. This method was compared to PBLHs obtained from potential temperature inversions and covariance wavelet transform retrievals from radiosondes and ceilometer aerosol backscatter profiles, respectively. A secondary thrust includes evaluating how convolutional autoencoders for denoising and in-painting could overcome deficiencies related to ceilometer signal noise and artifacts due to instrumentation issues and improve MLH detection. A third research effort is underway to evaluate how a Long Short Term Memory (LSTM) can be trained to learn how PBLH changes over time given geographical location, atmospheric conditions, and seasons. This effort evaluates how well the neural network can learn to predict how the PBLH changes over time. The results of these three efforts offer new ways to derive PBLH using machine learning that can result in improvements of future weather and air quality forecasting.