In the course of the 21st century, thawing of permafrost is expected to occur in large areas as a consequence of climate change, which could trigger a number of climatic feedback mechnisms on the local to global scale. As the vast and remote permafrost areas cannot be sufficiently covered by ground-based monitoring of soil temperatures in boreholes alone, it is desirable to exploit the wealth of multi-sensor-multi-source data to assess the thermal ground conditions on large scales. Soil temperatures can be modeled using Fourier’s law of heat conduction, so that the key challenge is to supply accurate time series of three key input variables at suitable spatial and temporal resolutions: 1. land surface temperature, 2. snow water equivalent, and 3. soil and snow thermal properties. In Norway, permafrost conditions range from mountain permafrost over organic-rich wetlands to high-arctic permafrost in Svalbard. Furthermore, the availability of gridded data sets from various sources makes it an well-suited test region to evaluate the performance of soil thermal models run with different input data, which facilitates comparing and benchmarking requirements on data quality. Surface temperatures are available from MODIS LST (1km resolution), interpolations from weather stations (1km resolution) and atmospheric modeling, e.g. reanalysis products (much coarser resolution). When computing weekly to monthly averages as needed for permafrost applications, all tested data sets display shortcomings. The overrepresentations of clear-sky conditions in MODIS LST can lead to a significant negative bias of wintertime snow surface temperatures, as demonstrated for a site on Svalbard. Interpolations from weather stations are problematic in mountain settings during stable atmospheric stratification conditions, with maximum biases found at mountain tops underlain by permafrost. We discuss data fusion approaches to achieve both quality assessment and improvement, with the goal to compile the best possible 50-year data set for permafrost applications. The snow water equilvalent is available from passive microwave remote sensing (25 km), interpolations from weather stations (1km resolution) and atmospheric modeling (much coarser resolution). None of these data sets can account for the considerable snow redistribution by wind on small scales. As a result, the snow depth is generally overestimated in permafrost mountain settings, which results in a warm-bias of modeled soil temperatures. Snow redistribution models based on the output of atmospheric modeling can provide probability density functions of snow depth, which allow a probabilistic assessment of permafrost conditions in larger grid cells. Finally, we highlight the importance of improved data sets on landcover and soil thermal properties, which currently constitute a major source of uncertainty for thermal permafrost modeling.