Planetary Boundary Layer Height Prediction over Karst Plateau: A Study Integrating Multi-Source Remote Sensing and Machine Learning
TL;DR
The results demonstrate that locally observation-driven machine learning can serve as a viable supplement or alternative to global reanalysis for boundary layer height estimation over complex karst terrain and they quantify the specific value that vertical profiling observations add relative to surface measurements alone.
Abstract
The planetary boundary layer height (PBLH) is a critical parameter for air quality modeling, weather forecasting, and climate studies, yet accurately estimating it over complex terrain remains a major challenge. This study investigates whether multi-source ground-based remote sensing combined with machine learning can provide reliable PBLH predictions in a karst plateau environment where the ERA5 reanalysis product shows large systematic errors. Using 17 months of continuous observations (from 1 January 2024 to 15 May 2025) from the Guiyang National Reference Climatological Station on the eastern Yunnan–Guizhou Plateau, including temperature and humidity profiles, wind profiles, and surface data, we constructed a 493-dimensional feature set and trained a CatBoost gradient boosting model. The model achieved a coefficient of determination (R2) of 0.80 and a root mean square error of 290 m on an independent test set, substantially outperforming ERA5, which yielded a negative R2 and a mean bias of approximately −500 m. Incorporating vertical profile features improved the R2 by 0.093 compared to using surface observations alone, with liquid water path, cloud base height, low-level temperature gradient, and low-level relative humidity emerging as the most important predictors. Classical physical diagnostic methods all produced negative R2 values when applied directly to the microwave radiometer retrieval data at this site, primarily due to fixed-segment artifacts across the seven height layers. When incorporated as additional prior features into CatBoost, none of the three configurations yielded meaningful improvement: R2 changed by at most +0.0010, RMSE by at most −0.7 m, and the 800 m classification accuracy by at most +0.61 percentage points, while the Bulk Richardson and Cn2 gradient priors slightly degraded RMSE. These results demonstrate that locally observation-driven machine learning can serve as a viable supplement or alternative to global reanalysis for boundary layer height estimation over complex karst terrain and they quantify the specific value that vertical profiling observations add relative to surface measurements alone. The findings provide valuable reference data for enhancing weather forecasting and air quality modeling, particularly in complex terrain areas.