以類神經網路探討雲林沿海地區地下水砷濃度與水質特徵

Artificial Neural Networks for Assessing Arsenic Concentrations and Characteristics of Groundwater Quality in the Coastal Area of Yun-Lin, Taiwan

林政華(Cheng-Hua Lin)

指導教授 ： 張斐章

國立臺灣大學/生物資源暨農學院/生物環境系統工程學研究所/碩士(2012年)

https://doi.org/10.6342/NTU.2012.00500

摘要

臺灣西南部沿海地區過去為烏腳病流行地區，經研究發現此情況與當地居民長期飲用含砷量高之地下水有密切關係，雖然當地居民已不再直接飲用地下水，但仍大量抽取地下水供給灌溉、養殖、公共及民生用水等多項用水標的，此舉可能造成砷透過生物累積過程進而對人體健康造成危害，實有必要建立可靠之推估模式，掌握該區域地下水砷污染情形，並分析該區域地下水之水質特性。
本研究目的為探討臺灣西南部沿海地區的水質特徵、影響地下水砷濃度消長之因子與可能影響砷釋出於地下水中的機制，並推估此區域地下水砷濃度之空間分布。本研究採用水利署於雲林縣沿海地區設置28座監測井之水質資料作為分析對象，首先將所有監測井水質資料輸入類神經網路-自組特徵映射網路(Self-Organizing Feature Map, SOM)進行聚類分析，經由視覺化結果呈現有效率地檢視水質因子間暨水質因子與砷濃度之關係，並由聚類的結果對測站進行分類以探討砷濃度在空間上的分布關係。進一步將自組特徵映射網路(SOM)的聚類結果移除砷的資訊作為決定輻狀基底函數類神經網路(Radial Basis Function Neural Network, RBFNN)之隱藏層神經元數目及中心點，再以RBFNN計算隱藏層至輸出層間神經元之權重值，以有限之水質因子推估砷濃度。上述推估結果亦與倒傳遞類神經網路(BPNN)模式的推估結果作比較，結論是SOM與RBFNN建立之模式精確度優於BPNN，且SOM於聚類時加入砷濃度資訊可使分類的結果更為顯著，並增加推估精確度；聚類分析的結果指出沿海地區測站地下水層鹽化與高砷污染情況，導致部分測站存在著高濃度砷，推論可能原因為此區域地下水環境趨於還原環境(如環境中存在高pH值、高鹼度、低溶氧量與低硫酸鹽濃度等)，有利於砷釋出，此結論與國內外相關研究結果一致。最後，應用地理資訊系統(GIS)將本模式推估之地下水砷濃度繪製地下水砷污染區域潛勢圖，呈現研究區域內砷濃度推估結果在1998年與1999年在空間上的變化，經由輸入水質因子便可了解此區域地下水中砷之釋出與遷移情形。
本模式特色在於可藉由SOM拓樸圖建構與解釋砷濃度與水質因子間關係，同時聚類結果有助於RBFNN網路之砷濃度推估能力，並以GIS呈現區域地下水砷濃度空間分布情形，展現以水質因子推估區域砷濃度分布圖，可供決策者透過視覺化了解區域時間與空間上的砷污染情形。

關鍵字

砷 ； 地下水水質 ； 類神經網路 ； 自組特徵映射網路 ； 輻狀基底函數類神經網路

並列摘要

In the past, Blackfoot disease commonly occurred along the southwestern coast of Taiwan. A number of investigations revealed this epidemic disease was highly related to arsenic (As) concentration in groundwater, which is the main source of drinking water to local residents. Although local residents do not directly drink groundwater any more in the Yun-Lin County of Taiwan, groundwater is still a main water source in this area because surface water suffers from limited sources. A large quantity of groundwater has been extracted from the aquifer for supplying water to the public, fish ponds and crop lands, which has resulted in the accumulation of arsenic in crops and fish. Products highly-contaminated with arsenic has threatened the health of residents. Therefore, it’s essential to construct a reliable model for estimating As concentration in groundwater.
The aims of this study are to assess the characteristics of groundwater quality, extract the factors affecting As concentration, investigate the sources releasing As, and estimate As concentration in groundwater. The Water Resources Agency (WRA) have set up 28 monitoring wells for investigating groundwater pollution, and water quality data collected by the WRA were used in this study. The first subject of this study is to import all the data collected form 28 wells to the Self-Organizing Feature Map (SOM) network. The SOM was applied to classifying all the water quality data into a topology map for finding the hidden relations among data and the spatial patterns between water quality variables and As concentration. Then the clustering results were adopted as the centroids of the Radial Basis Function Neural Networks (RBFNN) for accurately estimating As concentration based on water quality variables. In addition, the Back Propagation Neural Network (BPNN) was built to compare with the proposed model that integrates the SOM and RBF. The results demonstrate that the performance of the proposed model is better than the BPNN. When comparing the clustering results, adding As concentration to the SOM could make the clustering results more obvious and therefore achieves much accurate estimation. Moreover, the results demonstrate the characteristics of groundwater quality in coastal areas correlate with salinization and arsenic pollution factors. According to the clustering results, we surmise that the occurrence of high arsenic concentration in parts of wells is mainly because groundwater is in the reduction phase, especially at higher pH and Alk values and lower dissolved oxygen levels and SO42- concentration. Finally, the Geographic Information System (GIS) is applied to the results of groundwater quality models for displaying the spatial distribution map of As pollution so that we can realize the temporal and spatial variation in arsenic concentration in the study area.

並列關鍵字

Arsenic ； Groundwater quality ； Artificial neural network ； Self-Organizing Feature Map ； Radial Basis Function Neural Networks ； As ； ANN ； SOM ； RBFNN

參考文獻

43. 李品輝，2009，「以類神經網路探討全台蒸發量區域性分類與推估之成效」，國立臺灣大學生物環境系統工程學研究所碩士論文。

44. 高力山，2011，「人工智慧應用於區域地下水系統中砷污染推估之研究」，國立臺灣大學生物環境系統工程學研究所博士論文。

45. 高慧珊，2007，「以自組特徵映射網路推估蒸發量」，國立臺灣大學生物環境系統工程學研究所碩士論文。

48. 陳文福、呂學諭、劉聰桂，2010，「台灣地下水之氧化還原狀態與砷濃度」，農業工程學報，第56卷第2期。

52. 歐東坤，2005，「嘉南地區地下水砷濃度之研究」，國立臺灣大學生物環境系統工程學研究所碩士論文。