A data assimilation approach for simultaneously estimating a suite of land surface variables from satellite data

Abstract

After over two decade of efforts, many land products are now being produced systematically from a variety of satellite data, and these products have been widely used. However, estimating a set of atmospheric and surface variables from one sensor data is often an ill-posed inversion problem, because the number of unknowns is often larger than the available bands[1]. Thus, one has to make assumptions while trying to obtain realistic solutions, and as a result, most products still need significant improvements of quality and accuracy. Although the average accuracy may be acceptable, the error of each product can be very large under certain conditions. Furthermore, different products of land variables from different inversion algorithms are physically inconsistent for most cases. Many products in the current form are not suitable for climate study because the products are not continuous both spatially and temporally due to factors such as clouds. There is an urgent need to develop more advanced new inversion methods and produce more accurate products. We have recently proposed a data assimilation approach to estimate an improved suite of products from one or multiple satellite data. The general idea is to use the surface and atmospheric radiation models with parameters that are adjusted to optimally reproduce the spectral radiance received by the EOS sensors. Such adjustments are usually made by identifying reasonably close “first guesses” for the model parameters and determining statistically optimum estimates of the parameters by giving appropriate weights to the first guesses versus addition to the error increments needed to get agreement with the observations. The first guesses are the multiple years MODIS/MISR land product climatologies. The best estimate at present time is a climatological value corrected by some combinations of previous time's departure from climatology weighted using temporal autocorrelation and what it takes to fit present observations. The presentation will review this approach and also introduce three case studies[2-4]. Case one [3] estimated only leaf area index (LAI) by integrating temporal, spectral, and angular information from Moderate Resolution Imaging Spectroradiometer (MODIS), SPOT/VEGETATION, and Multi-angle Imaging Spectroradiometer (MISR) data based on an ensemble Kalman filter (EnKF) technique. Validation results at six sites demonstrate that the combination of temporal information from multiple sensors, spectral information provided by red and near-infrared (NIR) bands, and angular information from MISR bidirectional reflectance factor (BRF) data can provide a more accurate estimate of LAI than previously available. Case two [2] estimated temporally complete land-surface parameter profiles from MODIS time-series reflectance data also based on the EnKF technique. The products include LAI, the fraction of absorbed photosynthetically active radiation (FAPAR) and surface broadband albedo. The LAI/FAPAR and surface albedo values estimated using this framework were compared with MODIS collection 5 eight-day 1-km LAI/FAPAR products (MOD15A2) and 500-m surface albedo product (MCD43A3), and GEOV1 LAI/FAPAR products at 1/112. spatial resolution and a ten-day frequency, respectively, and validated by ground measurement data from several sites with different vegetation types. The results demonstrate that this new data assimilation framework can estimate temporally complete land-surface parameter profiles from MODIS time-series reflectance data even if some of the reflectance data are contaminated by residual cloud or are missing and that the retrieved LAI, FAPAR, and surface albedo values are physically consistent. The root mean square errors of the retrieved LAI, FAPAR, and surface albedo against ground measurements are 0.5791, 0.0453, and 0.0190, respectively. Case three [4] further estimated multiple land surface parameters and aerosol optical depth (AOD) from MODIS top-of-atmosphere (TOA) reflectance data without relying on atmospheric correction. Soil, vegetation canopy, and atmospheric radiative transfer models were coupled. LAI and AOD were estimated first and the coupled model then calculated land surface reflectance, incident photosynthetically active radiation (PAR), land surface albedo, and the FAPAR. The flowchart is shown in Fig. 1. The retrieved land surface parameters and AOD were compared with the corresponding MODIS, Global Land Surface Satellite (GLASS), GEOV1, and MISR products and validated by ground measurements from seven sites with different vegetation types. The results demonstrated that the new inversion method can effectively produce multiple physically consistent parameters with accuracy comparable to that of existing satellite products over the select sites (Figure 2).