Creating consistent datasets by combining remotely-sensed data and land surface model estimates through Bayesian uncertainty post-processing: The case of Land Surface Temperature from HIRS

Abstract

Abstract Satellite remote sensing allows many hydrological variables to be retrieved at regional and global scales, and these retrievals can contribute to our knowledge of these quantities and their spatial and temporal variation. However, the availability and precision of measurements from remote sensing are still limited. Other important sources of information are represented by reanalysis products and simulations from land surface models (LSMs), which are continuously improving their ability to reproduce the terrestrial hydrological and energy cycles. Such global datasets are very informative and they can be used to integrate the data provided by remote sensing. Among these datasets, the NCEP Climate Forecast System Reanalysis (CFSR) provides many hydrological variables covering the period from 1979 to 2011 with hourly temporal resolution and ~ 0.3-degree spatial resolution. This paper addresses the frequent challenge arising from remote sensing data: how can we create a long-term, global dataset starting with remote sensing data that are sparse in space and time, and provided by sensors flown on different satellites? This was the challenge for creating a long-term Land Surface Temperature (LST) dataset at an hourly resolution based on the High-Resolution Infrared Radiation Sounder (HIRS). To this end, a methodology is presented that merges the continuous CFSR LST estimates with the HIRS retrievals that come from 11 different NOAA satellites. The goals of the analysis are to have a long-term LST dataset consistent with the HIRS retrievals, and to assess the uncertainty associated with the reconstructed data through a Bayesian uncertainty post-processing methodology. The result of this approach is a temporally and spatially continuous dataset, with no step changes associated to the different NOAA satellites, and a dataset described in terms of a probability distribution function whose expected value is unbiased with respect to the original HIRS LST retrievals.