Estimation of Boundary Layer Height From Radar Wind Profiler by Deep Learning Algorithms

The boundary layer height (BLH) is a vital parameter that affects the vertical distribution of matter within the atmospheric boundary layer (ABL). However, the traditional algorithms determine the BLH based on changes in gradient within the signal-to-noise ratio (SNR) profile. It often leads to significant uncertainty under complex atmospheric conditions. Here, a convolutional neural network (CNN) algorithm considering multiple atmospheric parameter profiles is proposed for determining the BLH from radar wind profiler (RWP) data. The CNN algorithm is applied to the RWP dataset of atmospheric radiation measurement (ARM) site at Southern Great Plains (SGP) from August 2019 to July 2023. The sensitivity analysis shows that the CNN algorithm overcomes the shortcomings of the traditional algorithms that are susceptible to multiple local peaks. Moreover, the CNN algorithm performs well under complex conditions. It exhibits strong consistency with the BLH estimated by radiosonde (RS), with correlation coefficients, mean absolute error (MAE), and root-mean-square error (RMSE) of 0.81, 0.24, and 0.34 km, respectively. The CNN algorithm is then compared with the covariance wavelet transform (CWT) algorithm and the peak detection algorithm (PDA) using the BLH estimated by RS as a reference. The results indicate that the accuracy of BLH estimated by the CNN algorithm is higher than that of the two traditional algorithms. The MAE and RMSE of the CNN algorithm reduce from
$0.53~\pm ~0.56$
km (
$0.57~\pm ~0.60$
km) and 0.77 km (0.83 km) of CWT (PDA) to
$0.24~\pm ~0.25$
and 0.34 km, respectively. Finally, the diurnal and seasonal variation patterns of BLH are explored. The BLH shows a high correlation with solar radiation, rising from sunrise and then decreasing after sunset. Regarding seasonal variation, BLH peaks in summer and troughs in winter. Overall, the CNN algorithm proposed here can improve the accuracy and stability of BLH estimation. This study verifies the great potential of deep learning algorithms in the BLH estimation.