The global mean surface air temperature (SAT) has demonstrated the “unequivocal warming”. To understand the impact of the global warming, it is very important to quantify the spatial and temporal patterns of the surface air temperature change. Currently, most observational studies rely on in situ temperature measurements over the land and ocean. But the uneven and sparse nature of these temperature measurements may cause large uncertainty for the climate analysis especially at local and regional scales. With the rapid development of satellite data, it is possible to estimate spatial complete surface air temperature from satellite data using advanced statistical models. The satellite data-based estimation can serve as a better data source for local and regional climate analysis to reduce analysis uncertainty.

In this dissertation, I firstly examined the uncertainty of four mainstream gridded SAT datasets over the global land area (i.e., BEST-LAND, CRU-TEM4v, NASA-GISS, NOAA-NCEI). The comprehensive assessment of these datasets concludes that different data coverage may cause remarkable differences (i.e., -0.4 ~ 0.6°C) of calculated large scale (i.e., global, hemispheric) average SAT anomaly using different datasets. Moreover, these datasets show even larger differences at regional and local scale (5°×5°). The local and regional data differences can lead to statistically significant differences on linear trends of SAT estimated using different datasets. The correlation analysis shows strong relationship between the uncertainty of estimated SAT trends and the density of in situ measurements across different regions.

To reduce the uncertainty of surface air temperature data, I developed a statistical modelling framework which can estimate daily surface air temperature using remote sensing land surface temperature and radiation products. The framework uses machine learning models (i.e., rule-based Cubist regression model and multivariate adaptive regression spline) to characterize the physical difference between land surface temperature and surface air temperature by including radiation products at both surface and the top of the atmosphere. The model was firstly developed for the Tibetan Plateau using Cubist model trained with Chinese Meteorological Administration station measurements. Comprehensive evaluation show that the Cubist model can estimate the surface air temperature with nearly zero degree Celsius bias and small RMSEs between 1.6 °C ~ 2.1 °C. The estimated SAT over the entire Plateau for 2000-2015 show that the warming of the western part of the Plateau has been more prominent than the rest of the region. This result show the potential underestimation of conventional station measurements based studies because there are no station measurements to represent the rapid warming region.

The machine learning model is then extended to the northern high latitudes with necessary modification to account for the regional difference of the diurnal temperature cycle as well as the large data volume of the northern high latitudes. The MARS model trained using data over the northern high latitudes from the Global Historical Climatology Network daily data archive show a reasonable model performance with the bias of around -0.2 °C and the RMSE ranging between 2.1 – 2.6 °C. Further evaluation shows that the model performs worse over permanent snow and ice surface due to the insufficient training data to represent this specific surface conditions.

Overall, this research demonstrated that leveraging advanced statistical methods and satellite products can help generating high quality surface air temperature data which can provide much needed spatial details to reduce the uncertainty of local and regional climate analysis. The model developed in this research is generic and can be further extended to other regions with proper modification and training using high quality local data.