Developing a Land Continuous Variable Estimator to Generate Daily Land Products From Landsat Data

Generating spatially and temporally consistent biophysical products at the global scale from Landsat data for monitoring and assessing surface change dynamics remains a challenge. This article presents an inversion framework called Land continuous Variable Estimator (LoVE)–Landsat for estimating a group of spatiotemporal continuous land surface variables with daily temporal resolution from Landsat 5, 7, and 8 top-of-atmosphere (TOA) data. LoVE–Landsat adopts a data assimilation approach originally developed for coarse-resolution satellite data, such as Moderate Resolution Imaging Spectroradiometer (MODIS) and Visible Infrared Imaging Radiometer Suite (VIIRS). Major improvements to the approach include constructing empirical dynamic equations based on MODIS retrievals and other ancillary information, developing an artificial neural networks (ANN)-based emulator of the coupled radiative transfer (RT) model of atmosphere and land surface (vegetation, soil, and snow) as the observation operator, and designing a hybrid four-dimensional variational (4DVar) and ensemble Kalman filter (EnKF) data assimilation algorithm. The approach starts with generating the initial cloud-free regularly distributed (every 16 days) time series of Landsat data. The 4DVar is then used to assimilate clear-sky snow-free Landsat TOA observations over one year into the empirical dynamic evolution models of the land surface variables (e.g., leaf area index—LAI). The EnKF is then used to further adjust the state vector at the actual Landsat acquisition times. After determining a core set of variables (e.g., LAI), other variables, such as broadband albedo, emissivity, and fraction of absorbed photosynthetically active radiation (FAPAR), are calculated by the coupled RT model. Several experimental cases are presented to demonstrate that the proposed LoVE–Landsat framework is effective to estimate daily land surface variables.