Abstract：
The commonly used weather stations cannot fully capture the spatiotemporal variability of near-surface air temperature (Tair), leading to exposure misclassification and biased health effect estimates. We aimed to improve the spatiotemporal coverage of Tair data in Germany by using multi-stage modeling to estimate daily 1 × 1 km minimum (Tmin), mean (Tmean), maximum (Tmax) Tair and diurnal Tair range during 2000-2020. We used weather station Tair observations, satellite-based land surface temperature (LST), elevation, vegetation and various land use predictors. In the first stage, we built a linear mixed model with daily random intercepts and slopes for LST adjusted for several spatial predictors to estimate Tair from cells with both Tair and LST available. In the second stage, we used this model to predict Tair for cells with only LST available. In the third stage, we regressed the second stage predictions against interpolated Tair values to obtain Tair countrywide. All models achieved high accuracy (0.91 ≤ R2 ≤ 0.98) and low errors (1.03 °C ≤ Root Mean Square Error (RMSE) ≤ 2.02 °C). Validation with external data confirmed the good performance, locally, i.e., in Augsburg for all models (0.74 ≤ R2 ≤ 0.99, 0.87 °C ≤ RMSE ≤ 2.05 °C) and countrywide, for the Tmean model (0.71 ≤ R2 ≤ 0.99, 0.79 °C ≤ RMSE ≤ 1.19 °C). Annual Tmean averages ranged from 8.56 °C to 10.42 °C with the years beyond 2016 being constantly hotter than the 21-year average. The spatial variability within Germany exceeded 15 °C annually on average following patterns including mountains, rivers and urbanization. Using a case study, we showed that modeling leads to broader Tair variability representation for exposure assessment of participants in health cohorts. Our results indicate the proposed models as suitable for estimating nationwide Tair at high resolution. Our product is critical for temperature-based epidemiological studies and is also available for other research purposes.
摘要：
常用的气象站不能完全捕获近地表气温(Tair)的时空变化,导致暴露错误分类和有偏差的健康效应估计。我们旨在通过使用多阶段建模来估计每日1×1km最小值（Tmin）来改善德国Tair数据的时空覆盖，平均值（Tmean），2000-2020年期间的最大(Tmax)Tair和昼夜Tair范围。我们用了Tair气象站的观测,基于卫星的地表温度（LST），高程,植被和各种土地利用预测因子。在第一阶段,我们建立了一个线性混合模型，该模型具有每日随机截距和LST斜率，针对几个空间预测因子进行了调整，以从具有Tair和LST的细胞中估计Tair。在第二阶段,我们使用这个模型来预测只有LST可用的细胞的Tair。第三阶段,我们将第二阶段预测与插值Tair值进行回归，以获得Tair全国范围。所有模型均实现了高精度（0.91≤R2≤0.98）和低误差（1.03°C≤均方根误差（RMSE）≤2.02°C）。与外部数据的验证证实了良好的性能，本地,即，对于所有型号（0.74≤R2≤0.99，0.87°C≤RMSE≤2.05°C）和全国范围，对于Tmean模型（0.71≤R2≤0.99，0.79℃≤RMSE≤1.19℃）。年平均温度为8.56°C至10.42°C，2016年以后的年份比21年的平均值还要热。德国的空间变异性每年平均超过15°C，包括山脉，河流和城市化。使用案例研究，我们表明,建模可为健康队列参与者的暴露评估提供更广泛的Tair变异性表示.我们的结果表明，所提出的模型适用于高分辨率的全国Tair估计。我们的产品对于基于温度的流行病学研究至关重要，也可用于其他研究目的。
