Reconstruction of mesoscale precipitation fields from sparse observations in complex terrain

Abstract

The feasibility of a statistical reconstruction of mesoscale precipitation fields over complex topography from a sparse rain gauge network is examined. Reconstructions of gridded monthly precipitation for the European Alps (resolution 25 km, 1202 grid points) are derived from rain gauge samples (70-200-km interstation distance, 25-150 stations). The statistical model is calibrated over a 15-yr period, and the reconstructed fields are evaluated for the remaining 5 yr of the period 1971-90. The experiments are used to define the statistical setup, to assess the data requirements, and to describe the error statistics of a centennial reconstruction to be used in a forthcoming study. Reduced-space optimal interpolation is employed as the reconstruction method, involving data reduction by empirical orthogonal functions (EOFs) and least squares optimal estimation of EOF coefficients. Also, a procedure to define covariance-guided station samples with a "representative" spatial distribution for the reconstruction is proposed. Using a covariance-guided reference sample of 53 stations, the reconstruction accounts for 77% of the total variance. For individual grid points the relative reconstruction error (error variance divided by data variance) varies between 10% and 40%; this value drops to 2%-10% when considering subdomain means of 100 x 100 km2. The mesoscale patterns of the fields and multiyear precipitation anomalies are accurately reproduced. The EOF truncation is identified as the major limitation of the reconstruction skill but is necessary to avoid overfitting. Reconstructions from covariance-guided representative samples exhibit superior skill in comparison with those from randomly distributed stations. The skill of the reconstruction was found to depend marginally on the choice of the calibration period within the 20 yr, even when months with exclusively positive or negative values of the North Atlantic oscillation index were selected for calibration. This result indicates that the reconstruction model provides appreciable temporal stationarity.