|Title: An ensemble Kalman filter system with the Stony Brook Parallel Ocean Model v1.0 |
Authors: Ohishi et al.
I don’t think the authors have addressed my previous comments. To be scientifically sound and a valuable contribution for the ocean DA community, the manuscript definitely need further and detailed explanations for the results.
1. To my previous comment 1, “Compared to NOINFL, IAU in NOINFL+IAU degrades the accuracy. Why IAU degrade the accuracy for ocean assimilation that has longer time scale than atmosphere?” The authors explain that “The main difference between without and with the application of the IAU is directly updated the SSH or not. Temperature, salinity, horizontal velocities, and SSH analyses are used for the initial conditions for model integration within the assimilation window if the IAU method is not applied, whereas the analysis increments of temperature, salinity, horizontal velocities except for the SSH are distributed if the IAU method is applied.” I don’t understand why different update variables are used for experiments with and without IAU. Please give the detail configurations of the experiments in the text, and also explain the reasons for the different choices of updated variables for different assimilation experiments.
2. To my previous comment 2, “The authors state that IAU reduces the spread and accuracy of DA. But MULT, RTPP and RTPS have totally different impacts on the spread and accuracy when IAU is applied. Why MULT that also inflate the ensemble spread has the opposite impacts on spread and accuracy than RTPP and RTPS? Since the results with different inflation methods are inconsistent, it would be helpful to understand the roles of different inflation methods, especially the interactions with IAU.” The authors referenced an under-review manuscript that is not available for the reviewers. And the explanation is “The RTPP and RTPS relax the analysis ensemble perturbations toward the forecast ensemble perturbations. This implies that the analysis increments in the RTPP and RTPS would be smaller than the MULT, and the above degradation process might be suppressed.”
It would be helpful to see samples of analysis increments and subsequent forecasts, using MULT, RTPP and RTPS. However, the differences of analysis increments and subsequent forecasts with different inflation methods still cannot explain the interactions between IAU and inflation. Please provide understandings to the interactions of different inflation methods with IAU, which could help future design of data assimilation frameworks.
3. To my previous comment 3, “Previous studies of IAU (e.g., Lei and Whitaker 2016, He et al. 2020) showed that IAU has more advantages for variables that are more influenced by imbalances that variables that are less influenced by imbalances. However, results here are inconsistent with the previous findings. IAU improves the accuracy of wind field more than the accuracy of height field (Figures 3 and 4). Please provide explanations or insights for these counter-intuitive results.” The authors replied “The degradation of the accuracy by the IAU is consistent with He et al. (2020) who demonstrated that the accuracy of most variables is worser in the 3D-IAU experiment than experiment without the IAU when the assimilation windows are short of 1 and 3 hours [See table 3 of He et al. (2020)]; Lei and Whitaker (2016) who indicated that the accuracy of temperature and wind speed is worser in the 3D-IAU experiment than the experiment without the IAU using NCEP GFS experiments with assimilation of real observations [See fig. 8 of Lei and Whitaker (2016)]; and Yan et al. (2014) who showed that the IAU degrades the accuracy in twin experiments using an EnKF-based ocean data assimilation system [See table 3 of Yan et al. (2014)].”
First, Table 3 of He et al. (2020) showed that for the surface height that is more sensitive to imbalances than the other variables, 3DIAU is better than NoIAU for DA frequencies of 12h, 6h and 3h, while 3DIAU is worse than NoIAU for DA frequency of 1h. I don’t think the 1h DA frequency can be extrapolated here, since here less frequent observations are assimilated, and the oceanic model has much longer time scales than the QG model. Second, He et al. (2020) showed that IAU can impact the surface height more than the wind, since the latter is less sensitive to imbalances. But in this study, IAU improves the accuracy of wind field more than the accuracy of height field (Figures 3 and 4). Please provide dynamical explanations for this result.
4. To the last question of my previous comment 4, “The RMSD is computed for the prior or posterior? How the RMSD is computed for experiments with IAU?” and my previous comment 5, “Since assimilation is conducted at a daily frequency, both the daily prior and free forecast at longer forecast lead times worth to check.” The authors replied “To perform a free forecast after every assimilation cycle, all experiments must to be integrated again, and the huge amounts of the computational resources are required. Consequently, this is an issue in future studies.” I totally understand the computational cost. But since cycling assimilation experiments are already done, it should be straightforward to calculate the verifications with priors, since no additional computations needed. Moreover, just several samples of long free forecasts from different assimilation experiments could be useful to draw some conclusions.