Applying ANNs for Estimating the Regional Arsenic Pollution in Groundwater

本研究中應用類神經網路建立地下水砷濃度推估模式，以解決高度非線性砷汙染傳輸問題，及提高砷濃度推估之準確性。本研究以雲林縣沿海地區為研究區域，採用類神經網路建構地下水中砷推估模式，模式分為單一水井水質及區域水井水質之類神經網路模式，單一水井水質類神經網路模式主要針對單一監測井藉由砷與其他水質因子相關性，建立類神經網路推估模式，區域水井水質類神經網路模式則是應用全區資料，建立適用研究區域範圍之砷濃度推估模式，研究中除探討輸入因子及網路架構對模式誤差之影響外，並針對地下水水質模式較易遭遇到資料過少問題，提出交叉驗證法及修正型目標函數加以改善模式推估誤差，其中單一水井水質類神經網路模式之平均誤差 (rmse) 為 65.7 ug/l，區域水井水質類神經網路模式之平均誤差 (rmse) 為 112 ug/l，大部分監測井推估結果屬可接受範圍，僅#7監測井誤差較大，但對於本區砷變動較大且複雜地區，而監測資料有限下，藉由類神經網路達到可接受誤差，本模式成功解決過去傳統模式不易推估區域地下水中砷污染問題。最後，本研究將模式推估結果結合地理資訊系統 (GIS) 展示雲林縣沿海地區地下水中砷污染分布情形，可作為日後政府管理地下水之參考依據。
The groundwater extracted by some regional farmers leads to a lower level of the groundwater and a release of the poisonous substance in the groundwater. That affects the health of local residents, even the inhabitants who do not ingest the local agriculture and aquaculture products. The aim of the study is to build the arsenic water quality model by adopting the artificial neural networks (ANNs). Taking the YUN-LIN County as an example, the ANNs were constructed and assessed. The models are divided into two parts: (1) single well models and (2) regional well models. In the process of constructing the models, the optimal input factors and structures of the ANN models were discussed in this study. At the same time, we applied the cross validation method and the modified objective function to solving the data scarce problems of the monitoring wells. The results produced by the single well models and the regional well models were compared and demonstrated their applicability. Finally, the results obtained by the ANN models were integrated with GIS to display the distribution of the arsenic concentration at the coastal area in the YUN-LIN County. The results can offer a good reference to government decision-makers for the management of the groundwater and installation of monitoring wells.