Rising global temperatures pose a growing and urgent challenge, negatively impacting the life of billions of people due to heat-related issues, such as heightened mortality rates and decreased workforce productivity. To formulate effective countermeasures, policymakers require climate data with high-resolution in both spatial and temporal dimensions, coupled with high reliability, covering long-term trends and properly resolving extremes. Despite significant advancements of reanalysis datasets, contemporary models still lack kilometer-scale resolution and harbor inherent biases, compromising their suitability for informed decision-making. This thesis aims to address these limitations by proposing a methodology to enhance the spatial resolution and absolute accuracy of climate data. The approach involves a data fusion strategy, incorporating in-situ observations from automatic weather station networks, Remote Sensing data from diverse satellite missions, raw climate and reanalysis model inputs, and geospatial information. Several machine learning algorithms, including Decision Forests and Gradient Boosting, will be evaluated, and the most accurate model will be employed to generate a new global layer depicting air temperature. The model's performance will be assessed against weather station records and other high-resolution climate datasets to validate its accuracy. Additionally, an Urban Heat Island layer will be developed and compared with existing datasets, providing further insights into localized temperature variations in densely-populated areas.