Bathythermograph error analysis and reduction (BEAR)

Many oceanographic and tactical studies require high-fidelity sonar predictions, which require accurate portrayals of the ocean´s temperature structure. The Naval Oceanographic Office (NAVOCEANO) merges various oceanographic data into ´first-guess´ temperature fields. The Modular Ocean Data Assimilation System (MODAS) then assimilates more timely bathythermograph (BT) data to create temperature nowcasts and companion uncertainty fields. Cost constraints require minimization of at-sea measurements. The Sensor Placement for Optimal Temperature Sampling (SPOTS) algorithm determines the best placements of limited numbers of BTs to provide the most accurate temperature fields at the lowest possible cost. SPOTS hypothesizes that placements which minimize uncertainty will also tend to minimize error. The bathythermograph error analysis and reduction (BEAR) algorithm objectively determines covariance distances for use in MODAS assimilation routines and performs quality control (QC) on BTs to validate the SPOTS hypothesis. BEAR models physical errors that shift the temperature uniformly (factory mis-calibration, poor storage conditions, instrument abuse), shift the depth uniformly (starting lag and wave height), and expand or contract the gradients uniformly (inaccurate rate of fall). BEAR simulates the profiles that would result by backing out every possible combination of these physical errors. Deep-water temperature profiles sharply constrain BT QC, because they tend to be spatially and temporally stationary. An error of 0.4 degrees may correspond to fifty standard deviations from the mean. Unsurprisingly, the deeper half of the variances usually accounts for 95-99% of the total. Satellite-derived Sea-Surface Temperature (MCSST) data provide a second constraint, because their uncertainties are low (compared to BTs). BEAR computes the sum of variance over all depths for each simulated error-combination against these surface and climatological constraints. The minimum va- - riance sum is usually chosen to represent the most likely error combination, which is then applied to the BT to correct it, without at any point tampering with the temperature structure. Generally the one with the lowest variance is selected. However, if more than one BT is available from the same local region and within a reasonable time period, cross-correlation can be used to align the structural features (e.g., thermocline or mixed-layer depth) and select the best member of each variance cluster