Spatiotemporally resolved daily relative humidity predictions across Germany during 2000-2021: a Random Forest approach

Abstract

BACKGROUND AND AIM The lack of high-resolution relative humidity (RH) datasets and the limited ability of available weather stations to fully capture the spatiotemporal RH variability might lead to errors in exposure assessment for epidemiological studies and biased health effects estimates. We aimed to predict German-wide 1×1 km daily mean RH during 2000-2021 by using a machine learning approach. METHODS We used data from multiple sources, including RH observations, modelled precipitation and wind speed as well as remote sensing elevation, vegetation and the visible red, green and blue light bands. Our main predictor for estimating RH was modelled daily mean air temperature that we previously estimated in 1×1 km across Germany through a multi-stage regression-based approach incorporating two linear mixed models. Additionally, we included date in our RH model, capturing the day-to-day variation of the response-explanatory variables relationship. All the aforementioned predictors were included in a Random Forest (RF) model, which was separately developed for each year. RESULTS Ten-fold cross validation showed that the RF model achieved high accuracy (R²=0.81) and low errors (Root Mean Square Error (RMSE)=5.26%). We also compared our output with an independent dataset from a dense monitoring network in Augsburg metropolitan area (R²=0.81, RMSE=5.42%). Our models displayed high RH overall (21y-average RH=78.8%) and high spatial variability within a year across the country, exceeding 12.1% on average. RH distribution followed spatial patterns including urbanization, mountains, rivers and coastlines. For instance, the Alps and the North Sea coast were areas with elevated RH, while extended urban cores or individual cities were less humid than the surrounding rural settings. CONCLUSIONS Our findings indicate the proposed RF model as suitable for countrywide RH modeling at high spatiotemporal resolution, providing a reliable dataset for subsequent epidemiological analyses and other research purposes. KEYWORDS: relative humidity, spatiotemporal modeling, random forest, exposure assessment