Multisensor and multiscale data assimilation of remotely sensed

Multi-sensor and multi-scale data assimilation of remotely sensed snow observations Konstantinos Andreadis 1, Dennis Lettenmaier 1, and Dennis Mc. Laughlin 2 1. Department of Civil and Environmental Engineering, Box 352700, University of Washington, Seattle, WA 98195 2. Department of Civil and Environmental Engineering, Massachusetts Institute of Technology, Cambridge, MA 02139 Catchment-scale Hydrological Modeling & Data Assimilation International Workshop, 9 -11 January 2008, Melbourne, Australia ABSTRACT Forest Cover (%) A synthetic twin experiment is used to evaluate a data assimilation system that would ingest remotely sensed observations from passive microwave and visible wavelength sensors (snow water equivalent and snow cover extent derived products, respectively) with the objective of estimating snow water equivalent. Two data assimilation techniques are used, the Ensemble Kalman filter and the Ensemble Multiscale Kalman filter. One of the challenges inherent in such a data assimilation system is the discrepancy in spatial scales between the different types of snow-related observations. This study makes a first assessment of the feasibility of a system that would assimilate observations from multiple sensors and at different spatial scales for snow water equivalent estimation. Elevation (m) • Spatial maps of different SWE simulations for selected dates • Study domain is part of the upper Colorado River basin • Open-loop forcings created by perturbing precipitation and temperature with • Covers parts of Wyoming, Utah lognormal and gaussian multiplicative errors respectively and generating an and Colorado ensemble about those perturbed values • Relatively high elevation Truth Open-loop MSEn. KF Observed (average 2, 300 m) • Denser forest cover in SE, S and NW parts of the basin 1 Dec 2003 • Identical twin synthetic experiment • Importance of snow to the hydrologic cycle through its effects on water storage and land surface energy balance • Strategies for large scale observation of snow properties has focused on remote sensing • Visible wavelength sensors Snow Cover Extent observations No information on water storage and cloud cover limitations • Passive microwavelength sensors Brightness temperature a function of snow properties Snow water equivalent observations Problems with presence of wet snow, signal saturation and snow metamorphism • Additional information from hydrology models Forced with meteorological data and represent the effects of soils, topography and vegetation Uncertainties in forcing data and model parameters • Objective of study is to evaluate and compare data assimilation techniques using multi-scale remotely sensed observations of snow cover and water equivalent • Snow properties are simulated with the Variable Infiltration Capacity (VIC) model (Andreadis et al. , 2008) • Truth: model simulation with nominal forcings 15 Jan 2004 (precipitation and air temperature) • Open-loop: corrupt nominal forcings with errors, generate an ensemble about those, and simulate snow properties with that ensemble • Filter: model simulation with open-loop ensemble of forcings, and assimilation of synthetic observations (both En. KF and MSEn. KF) 10 Mar 2004 • Observations: synthetically generated by adding errors to true fields of snow water equivalent (SWE) and cover extent (SCE) • Spatial resolutions emulating MODIS aggregated to model resolution (~10 km) and AMSR-E (25 km) • Errors being N(0, 20 mm) for SWE and N(0, 0. 1) for SCE • Creating the tree topology • SCE observations on finest scale and SWE observations at scale immediately above, dictating tree levels to be 6 since finest scale is ~10 km • The algorithm moves from coarser scales down the tree, assigning blocks (no need to be rectangular) of model pixels to • Multiscale tree provides a physically consistent framework for assimilation of multi-sensor observations nodes based on a distance threshold first, and then elevation and forest cover as scale becomes finer • Similar performance between techniques, probably because of the • Tree topology represents the spatial structure of physiographic controls on snow accumulation and ablation processes selected tree topology (in order to have MODIS at finest scale and • Three criteria used to automatically assign states to neighboring nodes: distance, elevation, and forest cover • Two techniques are evaluated in this preliminary test, both based on the Ensemble Kalman filter (En. KF) • Model error covariance represented AMSR-E at one scale above) through an ensemble of model states • Increasing model spatial resolution (therefore increasing problem • Update occurs sequentially every time • Time series of basin-averaged SWE (left plot) and SCE (right plot) dimensionality), will lead to larger finest scale state vectors and an observation is available • Study period: 1 Sep 2003 – 31 May 2004 hypothetically larger differences between the En. KF and the • First technique: square root impleme- • MSEn. KF and En. KF simulations similar improvement over Open-loop MSEn. KF ntation of En. KF (Evensen, 2004) Basin Snow Water Equivalent Basin Snow Cover Extent • Perform similar experiment but assimilating passive and active • Second technique: Multiscale En. KF microwave brightness temperatures, and using a forward radiative (MSEn. KF, Zhou et al. 2007) transfer model (e. g. DMRT) • Covariances are represented through a multiscale tree that relates states through local parent-child relationships • States assigned to finest scale nodes, while measurements are assigned according to their spatial support Andreadis, K. , P. Storck, and D. Lettenmaier (2008): Modeling snow accumulation and ablation in forested environments, submitted to Water Resources Research. Evensen, G. (2004): Sampling strategies and square root analysis schemes for the En. KF, Ocean Dynamics, 54, 539 -560. Zhou, Y. , D. Mc. Laughlin, and D. Entekhabi (2007): An Ensemble multiscale filter for large nonlinear data assimilation problems, submitted to Monthly Weather Review.