Assessment of multi-sensor data assimilation for improved land surface model heat flux prediction

© 2014 Dr. Robert Carl PipunicThis thesis focuses on Land Surface Model (LSM) data assimilation for optimising latent (LE) and sensible (H) heat flux prediction. These fluxes influence cloud formation (hence precipitation) and atmospheric thermodynamics, therefore predicting them accurately is important for initialising Numerical Weather Prediction (NWP) forecasts. Applying data assimilation to update prognostic LSM state variables (typically soil moisture and soil temperature) with observed information can improve LE and H predictions. Furthermore, any overall improvement to modelled water balances will also have broader benefits in a water management context (i.e. for agriculture, water accounting etc.). Options for initialising LSM states includes assimilating screen-level atmospheric variables, and various remotely sensed land surface variables. Soil moisture and skin temperature data have been demonstrated as useful for directly impacting soil moisture and temperature states. The potential benefits of assimilating remotely sensed LE and H products into LSMs, for optimising LE and H prediction, have yet to be explored in detail. Evaluating LE and H assimilation is therefore a major objective of this thesis. It is examined in relation to assimilating different combinations of remote sensing data types, including soil moisture and skin temperature. Three main studies were undertaken: a synthetic-twin study; a one-dimensional study assimilating in-situ field observations; and a study assimilating remotely sensed data, including LE and H products. Field collection of LE and H (eddy covariance system), multi-depth soil moisture and temperature, and meteorological data over a full year was integral to this research. These data played a significant role in most experiments. The point scale proof-of-concept synthetic study demonstrated improved LE and H prediction from assimilating LE and H observations over a three month period, similar to improvements from skin temperature, and surpassing improvements made by soil moisture assimilation which produced the best root-zone soil moisture prediction. Assimilation frequency was important, with LE, H and skin temperature data representing cloud-free conditions, and approximating twice-daily MODIS thermal data, producing better fluxes than when representing the nearly fortnightly repeat time of Landsat. For the one-dimensional field data study, assimilation spanned one year with the time series of thermal related data filtered based on cloud-cover at the field site. Combined assimilation of different data types was also included. This study showed strong improvement to LE prediction from assimilating observed LE and H. While combined skin temperature and soil moisture assimilation improved LE the most, and also balanced this with simultaneous improvements to soil moisture and temperature states. Improved soil moisture prediction from assimilating only near-surface soil moisture data translated to smaller LE improvement than from assimilating LE and H, or combined skin temperature and soil moisture. The remotely sensed data study examined the assimilation of 25 km AMSR-E soil moisture data and of 5 km LE and H products derived from AQUA satellite based thermal observations. Experiments were performed with 5 km resolution model simulations spanning a year. Relative to data from 10 point scale in-situ stations, AMSR-E assimilation degraded the root-zone moisture predictions overall – with an average increase in RMSE of >100% compared to no assimilation. However, relative to single point-scale eddy covariance data, it improved LE prediction compared to no assimilation (RMSE reduced by 15%), with marginal improvement to H (reducing RMSE by ~1%). Assimilation of LE and H data derived from remote sensing improved both LE and H, reducing RMSE by ~13% and ~9% respectively compared to no assimilation. Repeated assimilation experiments using a range of increasing observational error for LE and H data (upwards from σ=50 Wm^-2) showed improvements to LE and H predictions with observational error of between 80 and 90 Wm^-2. Validating results from spatial remote sensing studies is limited with sparsely distributed in-situ point-scale data. Furthermore, scale discrepancies between assimilated remotely sensed data, simulation resolution and point validation data make it challenging to understand the cause of the apparent poorer moisture predictions from the AMSR-E assimilation. The most important finding from this thesis is that assimilating LE and H can improve LSM predictions of LE and H beyond improvements from soil moisture, and beyond improvements from assimilating skin temperature alone. Another key finding was the potential of combined near-surface soil moisture and skin temperature data assimilation to improve LE predictions beyond what assimilating LE and H data can, while also improving the root-zone state predictions. This is important to verify in follow-up studies given that remotely sensed LE and H require additional effort for derivation which is based on skin temperature data in the first place. The right combination of multiple data types may therefore be the best solution in the long run for optimising heat fluxes and improving water balance predictions