本研究以倒傳遞類神經網路(back-propagation neural network, BPN)結合半變異圖模式（semi-variogram model)及試誤法(try and error),建構一個地下水資料補遺模式，並以此模式對台南地區20個地 下水位測站進行月平均地下水位資料之補遺。研究結果顯示，資料若 先經過半變異圖模式所計算出之影響範圍(influence range)篩選後，的確 有助於BPN模式之地下水位補遺，但由於本研究地區之影響範圍並不 顯著，故本研究再利用試誤法篩選影響範圍内之地下水位資料。由研 究結果得知，經過試誤法篩選後，二十個測站補遺一個月至補遺三個 月之結果，與以自身測站資料作為BPN模式輸入項之模式相比，其效 率係數(coefficient of efficiency, CE)平均至少提高13%。以上。此外，輸入 BPN模式之相關測站數目也由原本平均8站減少到平均2站。根據上述 結果可知，經由半變異圖模式所計算出之影響範圍内之測站資料，若 再經由試誤法篩選後，不僅模式補遺結果之精確度明顯提升，且可大 幅地減少模式之資料輸入量及模式計算時間。此外，當地下水位之補 遺月數目越多時，本研究所提出之模式補遺結果表現越佳。

In this study, combination of the back-propagation neural network (BPN), semi-variogram model and try and error method to propose a groundwater level supplement model. The model is applied to actual groundwater data from 20 groundwater stations of the Tainan area. The study results show that data of selected stations within influence range obtained by semi-variogram model is indeed helpful to improve performance of BPN model in supplementing groundwater level. However, the influence range in the study area is unobvious. Try and error method is used to select groundwater level data within the influence range. The results show that there are 20 stations yield an average CE upgradation 13% at least. In addition, the average number of related stations is reduced eight to two. According to the above results, one can find that the proposed model can not only delete noise to the network and is more competent in supplementing groundwater level, but also can save training time of model and decrease costs of data collection and management. In addition, the proposed model produces a better performance with the long period supplement.