Stream water temperature in Austria – From irregular observations to regionalized monthly and daily datasets

Austria hosts a relatively dense network of stream water temperature stations (n = 680), most operating since 2006. Measurement frequency is highly irregular both across stations and within individual stations, ranging from 5-minute records to sporadic single observations with imprecise timestamps. These characteristics complicate analysis and modeling, and systematic assessments of trends and modeling approaches for stream water temperature remain scarce in Austria.

We present a workflow that addresses these obstacles to produce extended daily and monthly stream water temperature datasets for all available stations. Missing values are imputed on an hourly basis for days with at least one observation per station. We fit a station-specific diurnal spline weighted by daily meteorological predictors to reconstruct the diurnal cycle. Results are compared to an approach using LSTM autoencoder. From these reconstructions, we derive daily minimum, maximum, and mean temperatures together with day-specific uncertainty. Monthly statistics (e.g., quantiles, maximum, mean) are obtained via Monte Carlo simulation. To extend and regionalize the datasets, we use LSTMs for daily resolution and a combination of Model-based boosting and Topkriging for monthly estimates.

The resulting products enable robust trend analyses and the evaluation of stream water temperature models across Austria.