摘要本文研究的主要目的是評估海洋下表層各種變數對於中長期熱帶海溫(Sea Surface Temperature)的預報能力，進而尋找出能增進其預報技術的因子。我們使用的資料包括展期重建海溫資料庫(Extended Reconstructed Sea Surface Temperature)的熱帶海溫資料和全球海洋資料同化系統(Global Ocean Data Assimilation System)，簡易海洋資料同化系統(Simple Ocean Data Assimilation Reanalysis)，與海洋再分析資料系統(Ocean Reanalysis)等三種海洋資料庫的各個下表層(Below Sea Level)溫度，鹽度，和東西向流速變數。為了公平的評估海溫與下表層變數的預報能力，我們首先使用持續中和法去除掉海溫本身所具有的持續性。然後透過海溫與各個預變數之間的線性擬和所得的變異解釋百分比(explained variance percentage) 和相關係數對預報因子進行初步的篩選。最後再使用交叉驗證法(Cross-Validation)對這些篩選過後的預報因子進行更準確的預報技術評估。 研究結果顯示，線性擬合的變異解釋百分比與相關係數確實能幫助我們快速進行預報因子的篩選。但是，我們也發現這些結果中普遍出現深層變數有相當高的預報技術之情況。同時，交叉驗證的結果並未出現這種情況。因此，這種深層變數具有高預報技術的現象應該不是真實的。我們認為這是因為深層海洋觀測資料十分的缺乏，故在資料同化時必須大量人為的使用氣候平均補缺所產生的人工技術。本研究的結果顯示，我們仍然必須使用交叉驗證才能正確的挑選出真正有預報能力的預報因子。 經過比較三組資料庫中下表層變數交叉驗證的結果，海洋再分析資料系統資料庫下表層溫度第3與第4層(15m，25m)，以及下表層鹽度第2-8層(15m至90m) 在領先時間大於1年到18個月(lead time=12-18)的預報仍然有0.35以上的相關係數。這樣的預報技術明顯優於SST與其他深度層，也比另外兩組資料庫的變數所展現的預報能力高。透過比較三份資料庫同化方式上的不同，我們認為主要是因為海洋再分析資料系統引進了在等溫層同化鹽度的方式。除此之外，海洋再分析資料系統還比其他兩組資料庫在同化過程中多使用了偏差校正演算法(bias-correction algorithm)。我們懷疑這個演算法有可能造成資料中年循環的偏差，導致ORA-S3資料所預報的結果，在領先預報一年時有出現不尋常的高預報技術。AbstractThe capability of various oceanic sub-surface variables in predicting tropical sea surface temperature (SST) is examined in this study. The oceanic sub-surface variables include temperature, salinity, and zonal velocity at various depths from the Global Ocean Data Assimilation System (GODAS), the Simple Ocean Data Assimilation Reanalysis (SODA) and the Ocean reanalysis (ORA-S3). To prevent the distortion of forecast skill by persistence, the tropic SST was first normalized using the persistence neutralization transformation to neutralize its persistence. Then, the forecast capability was estimated using three measures. The first and the second measures were the percentages of explained variance and the correlation coefficients from linear fits between the tropic SSTA and a given subsurface variable. These two measures were used primarily for quick screening purpose. The third one was correlation coefficients from cross-validation procedure. This measure was used to estimate the true forecast skill of each examined predictor.Results from linear fits showed that percentages of explained variance and correlation coefficients can indeed speed up the screening of predictors. However, they also showed that variables at great depths had very high forecast skills, which were not reproduced in those of Cross-Validation. These artificial skills were suspected to be generated primarily due to the replacement of missing data with the corresponding climate mean in data assimilation stage. Therefore, to correctly select predictors with better forecast capability, the evaluation of forecast skill should be based on Cross-Validation results. Results from Cross-Validation showed that variables from ORA-S3 tended to have better skills than those from GODAS and SODA. However, ORA-S3 also showed unusual high forecast skills for lead times about 1yr. Compared assimilation procedure among these three datasets, it appears that ORA-S3 assimilates the salinity in the isothermal layer and incorporate a new bias correction algorithm in assimilation process may be the main reasons responsible for these phenomena.
