The hindcast results after the correction of UKF and two-SKF at non-measured stations are compared and summarized in Table 3, where the best estimates at each station are underlined. It can be seen that the two-SKF gives more accurate predictions of water level at points 2 and 3 comparing with the results of the UKF. This may be attributed to the fact that the process covariance which affects the filter efficiency is determined by two realizations in the two-SKF. As points 2 and 3 are located further away from the open boundary than points 1 and 4, their results computed by numerical model are less affected by the distortion at the open boundary. Therefore, the covariance calculated through realizations gives more promising results at these two points.
The statistical summaries of the distributed water level at non-measured points (CMB)
Station . | SRM . | UKF . | . | Two-SKF . | . |
---|---|---|---|---|---|
RMSE (m) . | RMSE (m) . | imp% . | RMSE (m) . | imp% . | |
1 | 0.4722 | 0.0657 | 86.1% | 0.0676 | 85.7% |
2 | 0.6619 | 0.0588 | 91.1% | 0.0293 | 95.6% |
3 | 0.6710 | 0.0597 | 91.1% | 0.0290 | 95.7% |
4 | 0.4558 | 0.0959 | 79.0% | 0.1029 | 77.4% |
Station . | SRM . | UKF . | . | Two-SKF . | . |
---|---|---|---|---|---|
RMSE (m) . | RMSE (m) . | imp% . | RMSE (m) . | imp% . | |
1 | 0.4722 | 0.0657 | 86.1% | 0.0676 | 85.7% |
2 | 0.6619 | 0.0588 | 91.1% | 0.0293 | 95.6% |
3 | 0.6710 | 0.0597 | 91.1% | 0.0290 | 95.7% |
4 | 0.4558 | 0.0959 | 79.0% | 0.1029 | 77.4% |
Underlined figures denote the best estimates for each method.