Assimilating stream flow, evapotranspiration and soil moisture data in AWRA-L model with particle filter

Abstract

The Australian Water Resource Assessment-Landscape model (AWRA-L) is calibrated against a selection of data from ~500 gauged catchments around Australia to identify a single optimised set of model parameter with an emphasis on improving streamflow prediction. However, this regional approach to AWRA-L calibration can lead to high uncertainty in estimation, especially in ungauged catchments. An approach to help improve prediction in ungauged area is the assimilation of remotely sensed data into hydrological models. Streamflow discharges (Q), satellite soil moisture (SM) and satellite evapotranspiration (ET) observations have been individually assimilated into hydrological models to improve predicted outputs. This paper aims to evaluate performance of both individual and joint assimilation of these three hydrological observations into the AWRA-L model using particle filter technique. The investigation used collected from six catchments across Australia with areas varying from 70 - 130 km2. In-situ streamflow data from the Hydrological Reference Stations (HRS) are divided by their respective catchment area to generate observations with units that are consistent with AWRA-L modelled streamflow. The European Space Agency Climate Change Initiative (ESA-CCI) soil moisture products were normalised to ensure that observed and simulated soil moisture are comparable. We disaggregated the 8-day CSIRO MODIS ReScaled potential ET (CMRS-ET) product to daily ET estimates using daily potential evapotranspiration (PET) and a linear interpolation method. Forcing data, initial conditions and spatial parameters of six catchments are collected from the datasets used for calibration and validation of the AWRA-L model for 2010. To address the limitation of high computational time required by the particle filter method in the grid-based AWRA-L model, we adopt a lump catchment approach where forcing inputs, initial conditions, and spatial parameters in each catchment were aggregated into one lumped value. Afterward the aggregated forcing is perturbed cell-wise using a normal distribution with 1000 samples to create a sample-based. These sample-based eventually are informed to the AWRA-L for the data assimilation process. Four assimilation scenarios were investigated, namely: (1) sole assimilation of Q, (2) sole assimilation of satellite SM, (3) sole assimilation of satellite ET, and (4) joint assimilation of Q, SM and ET. In addition, an open-loop simulation, i.e. without assimilation, was run as a reference. Statistical metrics such as correlation coefficient (R2), Nash-Sutcliffe model efficiency (NSE), root mean squared error (RMSE), and bias are used to assess the predicted outputs of the data assimilation model and open loop simulation. Initial results indicated that only assimilating Q was successful in improving ET predictions in all study catchments. The assimilation of ET, however, did not improve streamflow predictions. Although assimilation of soil moisture produces a slight improvement in ET prediction, it degrades the predicted streamflow in this study. Difference with the single assimilations, the joint assimilation of all three observations can improve all predictions compared with the open loop simulation in some catchments. In addition, we found that the model performs poorly in a catchment with very small streamflow. These findings suggest further research efforts in order to fully understand the data assimilation problem. Nevertheless, the results in this paper have demonstrated the potential uses of data assimilation to reduce prediction uncertainty of the AWRA-L model across Australia.