Using Gradient Boosting for gap-filling to analyze temperature and humidity patterns in an urban weather station network in Freiburg, Germany

Weather station networks (WSNs) are essential to investigate urban–rural and intra-urban variability of urban weather and climates. However, missing data in WSN time series lead to biases in the analysis of long-term WSN climatic statistics. Here, we use Extreme Gradient Boosting to impute gaps in observational air temperature and humidity time series of a WSN in an orographically complex, mid-sized European city (Freiburg, Germany). Imputation accuracy is evaluated across meteorological variables, stations, and artificially created gaps of 1–28 days. The gap-filling procedure shows good imputation accuracy with mean RMSEs of 0.46 K for air temperature and 2.51% for relative humidity. Model performance is insensitive to gap lengths but varies between stations, with larger errors in remote and non-built-up locations. Moreover, we use the gap-filled data to investigate urban–rural and intra-urban air temperature and humidity variability from September 2022 to August 2023. During the study period, the city center was 1.1 K warmer than representative rural areas, and vapor pressure was 7% lower. The annual number of tropical nights among stations in built-up locations varied between 0 and 29. Station clustering yields groups that are not only explainable by land cover and geographic proximity but also by orographic settings.