Sensible and Latent Heat Flux Prediction with Deep Learning using twenty years of Falkenberg micrometeorological data

The vertical fluxes of sensible and latent heat represent a major contribution to the exchange of energy between the land surface and the atmosphere. Their adequate description in numerical weather prediction and climate models is essential to realistically simulate near-surface weather conditions. Traditionally, these heat fluxes are parameterized relying on the Monin-Obukhov Similarity Theory (MOST) or the use of the Bulk-Richardson number. These parameterizations are based on differences in wind speed, air temperature, and humidity between adjacent measurement or model levels.

Wulfmeyer et al. (2023) estimated the heat fluxes with machine learning approaches and achieved a higher accuracy compared to MOST. Additionally, the analysis revealed radiation as a key predictor. However, their analysis is based on a rather short data period in August 2017 at three nearby locations in Oklahoma, USA, which limits the generalizability of the results.

In our study we replicate and expand the findings from Wulfmeyer et al. (2023) using a multilayer perceptron model (MLP) on a dataset from the boundary layer field site (GM) Falkenberg of the German Meteorological Service. The dataset consists of soil and meteorological variables over a period of twenty years, covering various seasons, synoptic weather situations and extreme weather events.

Our preliminary findings support the role of radiation as a dominant predictor for both the latent and sensible heat fluxes. We further studied the performance of the MLP for datasets of different lengths (e.g., one month as in Wulfmeyer et al., 2023, the same month over twenty years, or complete twenty-year data sets). Additionally, we tested the impact of removing redundancy in the selection of the predictor variables and the performance of the model under extreme conditions.