Evaluation of alternative model-data fusion approaches for retrospective water balance estimation

Abstract

Water resources observation and prediction systems are being developed in the Australian Bureau of Meteorology to produce water information services, and will include rolling water balance estimation. A prototype Australian Water Resources Assessment Model (AWRAM) has been developed, and the nationwide coverage, currency, accuracy, and consistency required means that remote sensing plays an important role. This paper tests and discusses alternative methods of blending models and observations. Integration of on-ground and remote sensing data into land surface models typically involves state updating through modeldata assimilation techniques. By comparison, retrospective water balance estimation and hydrological scenario modelling to date has mostly relied on non-sequential parameter estimation against stream flow observations, and has made little use of satellite earth observation. The most appropriate model-data fusion approach for a continental water balance estimation system will need to consider the trade-off between accuracy gains when using more sophisticated synthesis techniques and additional observations, and the computational overheads this incurs. This trade-off was investigated using relatively simple but wellperforming lumped models of seasonal vegetation dynamics and catchment hydrology that are implemented in the prototype AWRAM, while formal inter-comparison experiments to assess alternative component model paradigms and structures are underway. The performance of different model-data fusion (MDF) approaches was evaluated using flux tower ET measurements at four sites in Australia together with satellite observations of soil moisture over the corresponding area (AMSR-E passive microwave instrument). These observations, rather than hydrometric observations (e.g. streamflow), were chosen because of the more direct relationship they have with the site water balance over shorter time scales. Satellite-observed vegetation vigour (MODIS Enhanced Vegetation Index, EVI) was the assimilated variable. The MDF techniques tested include non-sequential estimation of model parameters (calibration against EVI, ET or both) and scaling of rainfall inputs, as well as sequential updating of leaf area index or soil moisture content using the ensemble Kalman filter. Non-sequential parameter estimation did not appear to provide much benefit compared to using prior parameter estimates, suggesting that the model parameterisation was comparatively robust and parameter values spatially invariant, at least when compared to errors in forcing data. A combination of parameter estimation and state updating did lead to improvements in some aspects of evaluation; reducing the apparent error in monthly evapotranspiration by 1% and in monthly top soil moisture content by 12%, respectively, when compared to using a priori parameter estimates. However it was also about three orders of magnitude more computationally intensive. Rainfall input adjustment was only tested in a relatively crude, non-sequential manner but results were encouraging, and appear to be a promising candidate for sequential approaches.