Intercomparison of Machine Learning Models to Determine the Planetary Boundary Layer Height Over Central Amazonia

Abstract

The planetary boundary layer height (zi) is a key parameter in meteorology and climatology, influencing weather prediction, cloud formation, and the vertical transport of scalars and energy near Earth's surface. This study compares multiple machine learning (ML) models that predict zi from surface measurements at two sites in Central Amazonia—the Amazon Tall Tower Observatory (ATTO) and the Manacapuru site of the GoAmazon experiment (T3). Models were trained on ceilometer data with radiosonde measurements used for validation. We evaluated model performance by withholding approximately 10% of the data (as complete months) for testing, comparing predictions against ERA-5 reanalysis data using RMSE, nRMSE, and R² metrics. Our results show that gradient boosted ensemble models using all available features perform best. A modified recursive feature elimination algorithm identified minimal sets of 5–7 surface measurements sufficient for accurate zi prediction, demonstrating potential for wider spatial monitoring using cost-effective sensors. The study revealed previously unrecognized variables influential in determining zi, such as deep soil temperature measurements (40 cm), suggesting new avenues for investigating land-atmosphere interactions. This study demonstrates the applicability of ML models to model zi.