Abstract:
We analyze the surface energy balance and the estimation of fluxes of heat and moisture from land to the atmosphere. We introduce a variational data assimilation scheme that use sequences of radiometric surface temperature measurements to estimate both surface boundary effects as well as moisture-related surface control on the partitioning between turbulent heat fluxes. The objective is to develop a methodology based on remote sensing data that will enable mapping of the energy balance components over extended areas. The surface moisture control on evaporation is captured by the dimensionless evaporative fraction (ratio of latent heat flux to the sum of the turbulent fluxes) which is nearly constant for near peak radiation hours on days without precipitation. The parameter capturing the turbulent transfer characteristics (bulk scalar turbulent transfer coefficient) includes the impacts of surface roughness effects and atmospheric stability conditions. The approach is tested over the FIFE field experiment site (Kansas, USA, 1987 and 1988). It is shown that sequential radiometric surface temperature data contain useful information on the partitioning of available surface energy and may even be used to infer some key characteristics of surface turbulent transfer. The feasibility of extending the land data assimilation to use only sparse samples of radiometric surface temperature measurements (corresponding to overpass times of currently operational satellites) is demonstrated through an observing system simulation experiment. The data assimilation model is applied to the Southern Great Plains 1997 (SGP97) field experiment, using land surface temperature maps obtained from three satellites, namely, Advanced Very High Resolution Radiometer (AVHRR), the Special Sensor Microwave/Imager (SSM/I) and Geostationary Operational Environment Satellite (GOES). The application over the SGP domain gives reasonable estimates of surface fluxes, evaporative fraction and roughness-related parameters, confirming the potential of the land data assimilation scheme as an operational tool for the monitoring of surface energy balance and fluxes.

Abstract:
Most data assimilation algorithms are based on classical statistical methods that adopt pixel-oriented descriptions of natural phenomena. While these methods are convenient they can have difficulty capturing the distinctive spatial and temporal features of special interest in geoscience applications. Such features include storm systems, ocean currents, algae blooms, geological formations, and wildfires. The problem is that the spatial structure that is clearly evident in nature may not be adequately conveyed through the means and covariances of large pixel-oriented state vectors. In this talk we consider how methods from computer vision and machine learning might be applied to feature-based data assimilation problems. These methods typically adopt a Bayesian perspective that focuses on the population of possible features rather than the population of possible state vectors. The objective is to derive an ensemble of conditional features that are samples from the posterior Bayesian distribution (i.e. incorporate both observations and prior information). Prior information is conveyed by an ensemble of prior features generated from training images. These images may be postulated, derived from observations, or obtained from physically-based stochastic models. Posterior features can be generated with Monte Carlo sampling techniques that make relatively few assumptions but are currently computationally infeasible for large problems. The success of a feature-based data assimilation will depend on our ability to 1) characterize the distinctive properties of natural features in feature spaces that are significantly smaller than the original pixel-based space (e.g. using image compression techniques), 2) generate candidate samples (proposals) that incorporate measurement information while remaining physically realistic (i.e. consistent with the training image), 3) determine the prior probability and likelihood of candidate features (i.e. define probability measures over feature spaces) , 4) develop computationally efficient reduced-order models that make it possible to generate and test a large number of candidate features. In this talk we illustrate concepts and assess some of the major research challenges and opportunities posed by a feature-based perspective.

Abstract:
The variational data assimilation problem can be solved in either its primal (3D/4D-Var) or dual form (3D/4D-PSAS). As shown in El Akkraoui et al. (2008), both methods are equivalent at convergence but the dual method exhibits a spurious behavior at the beginning of the minimization which leads to less probable states than the background state. This is a serious concern when using the dual method in operational implementations when only a finite number of iterations can be afforded. Two minimization algorithms are examined: the Conjugate Gradient (CG) and the Minimum Residual (Minres) methods. The Minres algorithm insures a monotonic decrease of the gradient norm and when applied to the dual problem, it also leads to a monotonic decrease of the primal cost function. Moreover, it is shown that a new termination criteria, based on the error norm in model space, can be used in the dual case to achieve the same accuracy in the analysis state when only a finite number of iterations are completed. This is of great importance for an implementation of the dual form of a weak-constraint 4D-Var. This will be discussed in the presentation and preliminary results will be presented.

Abstract:
The Atmospheric Infrared Sounder (AIRS) onboard the NASA Aqua satellite and the Infrared Atmospheric Sounding Interferometer (IASI) onboard the METOP-A satellite both are cutting-edge spectrometers for 3-dimensional mapping of air and surface temperature, moisture, greenhouse gases, and cloud properties. With several kilo-channel measurements in the infrared spectral coverage, AIRS and IASI has a spectral resolution more than 100 times greater than previous IR sounders, and provides more accurate information to improve weather forecasting and support climate research. To enjoy the greatest advantage of such high resolution infrared observations, high-performance-computing radiative transfer models for AIRS and IASI are needed to facilitate more effective applications in physical retrievals and data assimilation. Approximately every six months there is a doubling in the speed of Graphics Processing Units (GPUs). Currently, the flagship NVIDIA GPU has 240 computing cores, compared to the best INTEL CPU with just 6 cores. The computing performance of the GPU has significantly outpaced its CPU counterpart, with a theoretical peak performance of 1 TFlops per GPU in single precision and 345 GFlops per GPU in double precision. The combined features of general-purpose supercomputing, high parallelism, high memory bandwidth (102 GB per GPU), low cost, and compact size are what make a GPU-based desktop computer an appealing alternative to a massively parallel system made up of commodity CPUs (e.g. Beowulf clusters). This paper presents our effort to develop the GPU-based high-performance AIRS/IASI radiative transfer model running on NVIDIA GPUs via CUDA (Compute Unified Device Architecture), the compute engine in NVIDIA GPUs for massively multi-threaded parallel computation. Our GPU implementation of the forward model is tested on a low-cost (~$8,000) NVIDIA S1070 personal supercomputer with 4 Tesla GPUs (total 960 cores) delivering 4 TFlops peak performance. The result is compared with the native INTEL multi-core CPU implementation to show the significant speed-ups of computing the AIRS/IASI radiance spectrum using a GPU-based system. This is our first step towards the development of a GPU-based high-performance full-spectrum AIRS/IASI physical retrieval system for product applications and data assimilation.

Abstract:

Abstract:

Abstract:

Abstract:

Abstract:

Abstract: