Adaptive Observation Strategies with the Local Ensemble Transform Kalman Filter

Adaptive observation strategies (AOS) aim to improve forecasts by adding additional observations at a few locations that have no standard observations. Lorenz and Emanuel (1998) designed experiments to evaluate different adaptive strategies with Lorenz 40-variable model. Routine observations are observed over “land” (grid points from 21 to 40) every 6 hours. One adaptive point is chosen from one of the points over “ocean” (grid points from 1 to 20) every 6 hours. They found that the performance of adaptive methods (multiple breeding, multiple replication, singular vector) is better than random choice. The best result was obtained from multiple replication (a variation of multiple breeding with perturbed observations). With a 1024 ensemble members Ensemble Kalman Filter (EnKF) assimilation scheme, Hansen and Smith (2000) got comparable results from singular vector adaptive observation strategy (SVAOS) as with the other methods investigated by Lorenz and Emanuel (1998), who had concluded that SVAOS is inferior to the other methods. Trevisan and Uboldi (2004) used the most unstable vector of the observation-analysis-forecast (OAF) system (obtained by breeding) to choose the adaptive point and to define the analysis increment. Their average analysis errors over the ocean were reduced by 50% compared with Lorenz and Emanuel (1998). Here we explore the use of an optimal adaptive method with the Local Ensemble Transform Kalman Filter (LETKF, Hunt, 2005). LETKF is a square root EnKF that does data assimilation over small local patches around each grid point and is thus very parallel and efficient. In these experiments we use only 15 ensemble members, and use different measures of ensemble uncertainty to choose the point of maximum uncertainty where the next adaptive observation should be made. The strategies include ensemble-spread method, local a P method and a combined method. All three methods result in similar results that are better than previous results. Ensemble-spread method is almost computational free, but it doesn’t consider the possible effect of the future adaptive observation error. Local a P method considers the effects of the adaptive observation error, but it is very expensive. The combined method combines the advantages of both methods.